Implement multiqueue virtio-net
Subsystems (6)
v2 cover Implement multiqueue virtio-net
Following patches implement transmit MQ in virtio-net. Also included is the user qemu changes. MQ is disabled by default unless qemu specifies it.
1. This feature was first implemented with a single vhost. Testing showed 3-8% performance gain for upto 8 netperf sessions (and sometimes 16), but BW dropped with more sessions. However, adding more vhosts improved BW significantly all the way to 128 sessions. Multiple vhost is implemented in-kernel by passing an argument to SET_OWNER (retaining backward compatibility). The vhost patch adds 173 source lines (incl comments).
2. BW -> CPU/SD equation: Average TCP performance increased 23% compared to almost 70% for earlier patch (with unrestricted #vhosts). SD improved -4.2% while it had increased 55% for the earlier patch. Increasing #vhosts has it's pros and cons, but this patch lays emphasis on reducing CPU utilization. Another option could be a tunable to select number of vhosts threads.
3. Interoperability: Many combinations, but not all, of qemu, host, guest tested together. Tested with multiple i/f's on guest, with both mq=on/off, vhost=on/off, etc.
Changes from rev1:
------------------
1. Move queue_index from virtio_pci_vq_info to virtqueue, and resulting changes to existing code and to the patch.
2. virtio-net probe uses virtio_config_val.
3. Remove constants: VIRTIO_MAX_TXQS, MAX_VQS, all arrays allocated on stack, etc.
4. Restrict number of vhost threads to 2 - I get much better cpu/sd results (without any tuning) with low number of vhost threads. Higher vhosts gives better average BW performance (from average of 45%), but SD increases significantly (90%).
5. Working of vhost threads changes, eg for numtxqs=4:
vhost-0: handles RX
vhost-1: handles TX[0]
vhost-0: handles TX[1]
vhost-1: handles TX[2]
vhost-0: handles TX[3]
Enabling MQ on virtio:
-----------------------When following options are passed to qemu:
- smp > 1
- vhost=on
- mq=on (new option, default:off)
then #txqueues = #cpus. The #txqueues can be changed by using an optional 'numtxqs' option. e.g. for a smp=4 guest:
vhost=on -> #txqueues = 1
vhost=on,mq=on -> #txqueues = 4
vhost=on,mq=on,numtxqs=8 -> #txqueues = 8
vhost=on,mq=on,numtxqs=2 -> #txqueues = 2 Performance (guest -> local host):
-----------------------------------System configuration:
Host: 8 Intel Xeon, 8 GB memory
Guest: 4 cpus, 2 GB memory, numtxqs=4
All testing without any system tuning, and default netperf Results split across two tables to show SD and CPU usage:
________________________________________________________________________
TCP: BW vs CPU/Remote CPU utilization:
# BW1 BW2 (%) CPU1 CPU2 (%) RCPU1 RCPU2 (%)
________________________________________________________________________
1 69971 65376 (-6.56) 134 170 (26.86) 322 376 (16.77) 2 20911 24839 (18.78) 107 139 (29.90) 217 264 (21.65) 4 21431 28912 (34.90) 213 318 (49.29) 444 541 (21.84) 8 21857 34592 (58.26) 444 859 (93.46) 901 1247 (38.40) 16 22368 33083 (47.90) 899 1523 (69.41) 1813 2410 (32.92) 24 22556 32578 (44.43) 1347 2249 (66.96) 2712 3606 (32.96) 32 22727 30923 (36.06) 1806 2506 (38.75) 3622 3952 (9.11) 40 23054 29334 (27.24) 2319 2872 (23.84) 4544 4551 (.15) 48 23006 28800 (25.18) 2827 2990 (5.76) 5465 4718 (-13.66) 64 23411 27661 (18.15) 3708 3306 (-10.84) 7231 5218 (-27.83) 80 23175 27141 (17.11) 4796 4509 (-5.98) 9152 7182 (-21.52) 96 23337 26759 (14.66) 5603 4543 (-18.91) 10890 7162 (-34.23) 128 22726 28339 (24.69) 7559 6395 (-15.39) 14600 10169 (-30.34)
________________________________________________________________________
Summary: BW: 22.8% CPU: 1.9% RCPU: -17.0%
________________________________________________________________________
TCP: BW vs SD/Remote SD:
# BW1 BW2 (%) SD1 SD2 (%) RSD1 RSD2 (%)
________________________________________________________________________
1 69971 65376 (-6.56) 4 6 (50.00) 21 26 (23.80) 2 20911 24839 (18.78) 6 7 (16.66) 27 28 (3.70) 4 21431 28912 (34.90) 26 31 (19.23) 108 111 (2.77) 8 21857 34592 (58.26) 106 135 (27.35) 432 393 (-9.02) 16 22368 33083 (47.90) 431 577 (33.87) 1742 1828 (4.93) 24 22556 32578 (44.43) 972 1393 (43.31) 3915 4479 (14.40) 32 22727 30923 (36.06) 1723 2165 (25.65) 6908 6842 (-.95) 40 23054 29334 (27.24) 2774 2761 (-.46) 10874 8764 (-19.40) 48 23006 28800 (25.18) 4126 3847 (-6.76) 15953 12172 (-23.70) 64 23411 27661 (18.15) 7216 6035 (-16.36) 28146 19078 (-32.21) 80 23175 27141 (17.11) 11729 12454 (6.18) 44765 39750 (-11.20) 96 23337 26759 (14.66) 16745 15905 (-5.01) 65099 50261 (-22.79) 128 22726 28339 (24.69) 30571 27893 (-8.76) 118089 89994 (-23.79)
________________________________________________________________________
Summary: BW: 22.8% SD: -4.21% RSD: -21.06%
________________________________________________________________________
UDP: BW vs SD/CPU
# BW1 BW2 (%) CPU1 CPU2 (%) SD1 SD2 (%)
_____________________________________________________________________________
1 36521 37415 (2.44) 61 61 (0) 2 2 (0) 4 28585 46903 (64.08) 397 546 (37.53) 72 68 (-5.55) 8 26649 44694 (67.71) 851 1243 (46.06) 334 339 (1.49) 16 25905 43385 (67.47) 1740 2631 (51.20) 1409 1572 (11.56) 32 24980 40448 (61.92) 3502 5360 (53.05) 5881 6401 (8.84) 48 27439 39451 (43.77) 5410 8324 (53.86) 12475 14855 (19.07) 64 25682 39915 (55.42) 7165 10825 (51.08) 23404 25982 (11.01) 96 26205 40190 (53.36) 10855 16283 (50.00) 52124 75014 (43.91) 128 25741 40252 (56.37) 14448 22186 (53.55) 133922 96843 (-27.68)
____________________________________________________________________________
Summary: BW: 50.4 CPU: 51.8 SD: -27.68
_____________________________________________________________________________
N#: Number of netperf sessions, 60 sec runs BW1,SD1,RSD1: Bandwidth (sum across 2 runs in mbps), SD and Remote
SD for original code
BW2,SD2,RSD2: Bandwidth (sum across 2 runs in mbps), SD and Remote
SD for new code.
CPU1,CPU2,RCPU1,RCPU2: Similar to SD.
For 1 TCP netperf, I ran 7 iterations and summed it. Explanation for degradation for 1 stream case:
1. Without any tuning, BW falls -6.5%.
2. When vhosts on server were bound to CPU0, BW was as good as with original code.
3. When new code was started with numtxqs=1 (or mq=off, which is the default), there was no degradation.
Next steps:
-----------
1. MQ RX patch is also complete - plan to submit once TX is OK (as well as after identifying bandwidth degradations for some test cases).
2. Cache-align data structures: I didn't see any BW/SD improvement after making the sq's (and similarly for vhost) cache-aligned statically:
struct virtnet_info {
...
struct send_queue sq[16] ____cacheline_aligned_in_smp;
...
};3. Migration is not tested.
Review/feedback appreciated.
---
RFC PATCH v2 1/4 Change virtqueue structure
Move queue_index from virtio_pci_vq_info to virtqueue. This allows callback handlers to figure out the queue number for the vq that needs attention.
--- drivers/virtio/virtio_pci.c | 10 +++------- include/linux/virtio.h | 1 + 2 files changed, 4 insertions(+), 7 deletions(-)
include/linux/virtio.h +1 -0
@@ -22,6 +22,7 @@ struct virtqueue {void (*callback)(struct virtqueue *vq);const char *name;struct virtio_device *vdev; +int queue_index;/* the index of the queue */ void *priv; };
drivers/virtio/virtio_pci.c +3 -7
@@ -75,9 +75,6 @@ struct virtio_pci_vq_info/* the number of entries in the queue */ int num; -/* the index of the queue */ -int queue_index; -/* the virtual address of the ring queue */ void *queue;
@@ -185,11 +182,10 @@ static void vp_reset(struct virtio_devicstatic void vp_notify (struct virtqueue *vq) {struct virtio_pci_device *vp_dev =to_vp_device (vq->vdev); -struct virtio_pci_vq_info *info = vq->priv;/* we write the queue's selector into the notification register to * signal the other end */ -iowrite16 (info->queue_index, vp_dev->ioaddr +VIRTIO_PCI_QUEUE_NOTIFY ); +iowrite16 (vq->queue_index, vp_dev->ioaddr +VIRTIO_PCI_QUEUE_NOTIFY ); }/* Handle a configuration change: Tell driver if it wants to know. */
@@ -385,7 +381,6 @@ static struct virtqueue *setup_vq(structif (!info)return ERR_PTR (-ENOMEM ); - info->queue_index = index; info->num = num; info->msix_vector = msix_vec;
@@ -408,6 +403,7 @@ static struct virtqueue *setup_vq(struct goto out_activate_queue; } + vq->queue_index = index; vq->priv = info; info->vq = vq;
@@ -446,7 +442,7 @@ static void vp_del_vq(struct virtqueue *list_del (&info->node);spin_unlock_irqrestore (&vp_dev->lock, flags); -iowrite16 (info->queue_index, vp_dev->ioaddr +VIRTIO_PCI_QUEUE_SEL ); +iowrite16 (vq->queue_index, vp_dev->ioaddr +VIRTIO_PCI_QUEUE_SEL );if (vp_dev->msix_enabled) { iowrite16(VIRTIO_MSI_NO_VECTOR,
RFC PATCH v2 2/4 Changes for virtio-net
In reply to Krishna KumarImplement mq virtio-net driver.
Though struct virtio_net_config changes, it works with old qemu's since the last element is not accessed, unless qemu sets VIRTIO_NET_F_NUMTXQS.
--- drivers/net/virtio_net.c | 213 ++++++++++++++++++++++++++--------- include/linux/virtio_net.h | 3 2 files changed, 163 insertions(+), 53 deletions(-)
include/linux/virtio_net.h +3 -0
@@ -26,6 +26,7 @@#define VIRTIO_NET_F_CTRL_RX 18/* Control channel RX mode support */ #define VIRTIO_NET_F_CTRL_VLAN 19/* Control channel VLAN filtering */ #define VIRTIO_NET_F_CTRL_RX_EXTRA 20/* Extra RX mode control support */ +#define VIRTIO_NET_F_NUMTXQS 21/* Device supports multiple TX queue */ #define VIRTIO_NET_S_LINK_UP 1/* Link is up */
@@ -34,6 +35,8 @@ struct virtio_net_config {__u8 mac[6 ];/* See VIRTIO_NET_F_STATUS and VIRTIO_NET_S_* above */ __u16 status; +/* number of transmit queues */ +__u16 numtxqs; } __attribute__((packed));/* This is the first element of the scatter-gather list. If you don't
drivers/net/virtio_net.c +160 -53
356 lines not on this page · raw message
Re: [v2 RFC PATCH 2/4] Changes for virtio-net
Quoted 36 lines · Krishna Kumar
Implement mq virtio-net driver.
Though struct
virtio_net_configchanges, it works with old qemu's since the last element is not accessed, unless qemu setsVIRTIO_NET_F_NUMTXQS.Signed-off-by Krishna Kumar <[email protected]>--- drivers/net/virtio_net.c | 213 ++++++++++++++++++++++++++--------- include/linux/virtio_net.h | 3 2 files changed, 163 insertions(+), 53 deletions(-)diff -ruNp
org2/include/linux/virtio_net.htx_only2/include/linux/virtio_net.hinclude/linux/virtio_net.h +3
-0@@ -26,6 +26,7 @@#define VIRTIO_NET_F_CTRL_RX 18/* Control channel RX mode support */ #define VIRTIO_NET_F_CTRL_VLAN 19/* Control channel VLAN filtering */ #define VIRTIO_NET_F_CTRL_RX_EXTRA 20/* Extra RX mode control support */ +#define VIRTIO_NET_F_NUMTXQS 21/* Device supports multiple TX queue */ #define VIRTIO_NET_S_LINK_UP 1/* Link is up */ @@ -34,6 +35,8 @@ struct virtio_net_config {__u8 mac[6 ];/* See VIRTIO_NET_F_STATUS and VIRTIO_NET_S_* above */ __u16 status; +/* number of transmit queues */ +__u16 numtxqs; } __attribute__((packed));/* This is the first element of the scatter-gather list. If you don't diff -ruNp
org2/drivers/net/virtio_net.ctx_only2/drivers/net/virtio_net.c
drivers/net/virtio_net.c module_param(gso, bool, 0444); · line 49
Quoted 10 lines
drivers/net/virtio_net.c +7
-0@@ -40,9 +40,20 @@ module_param(gso, bool, 0444);#define VIRTNET_SEND_COMMAND_SG_MAX 2 +/* Our representation of a send virtqueue */ +struct send_queue { +struct virtqueue *svq; + +/* TX: fragments + linear part + virtio header */ +struct scatterlist tx_sg[MAX_SKB_FRAGS +2 ]; +};
You probably want ____cacheline_aligned_in_smp
drivers/net/virtio_net.c module_param(gso, bool, 0444); · line 57
Quoted 9 lines
drivers/net/virtio_net.c +5
-1+struct virtnet_info {struct virtio_device *vdev; -struct virtqueue *rvq, *svq, *cvq; +int numtxqs;/* Number of tx queues */ +struct send_queue *sq; +struct virtqueue *rvq; +struct virtqueue *cvq;struct net_device *dev;
struct napi will probably be dirtied by RX processing
You should make sure it doesnt dirty cache line of above (read mostly) fields
Quoted 148 lines
drivers/net/virtio_net.c +28
-23struct napi_struct napi;unsigned int status;@@ -62,9 +73,8 @@ struct virtnet_info {/* Chain pages by the private ptr. */ struct page *pages; -/* fragments + linear part + virtio header */ +/* RX: fragments + linear part + virtio header */ struct scatterlist rx_sg[MAX_SKB_FRAGS +2 ]; -struct scatterlist tx_sg[MAX_SKB_FRAGS +2 ]; };struct skb_vnet_hdr {@@ -120,12 +130,13 @@ static struct page *get_a_page(struct vistatic void skb_xmit_done (struct virtqueue *svq) {struct virtnet_info *vi = svq->vdev->priv; +int qnum = svq->queue_index -1 ;/* 0 is RX vq */ /* Suppress further interrupts. */ virtqueue_disable_cb (svq);/* We were probably waiting for more output buffers. */ -netif_wake_queue (vi->dev); +netif_wake_subqueue (vi->dev, qnum); } static void set_skb_frag(struct sk_buff *skb, struct page *page,@@ -495,12 +506,13 @@ again:return received; } -static unsigned int free_old_xmit_skbs (struct virtnet_info *vi) +static unsigned int free_old_xmit_skbs (struct virtnet_info *vi, +struct virtqueue *svq) {struct sk_buff *skb;unsigned int len, tot_sgs =0 ; -while ((skb =virtqueue_get_buf (vi->svq, &len)) !=NULL ) { +while ((skb =virtqueue_get_buf (svq, &len)) !=NULL ) {pr_debug ("Sent skb %p\n" , skb); vi->dev->stats.tx_bytes += skb->len; vi->dev->stats.tx_packets++;@@ -510,7 +522,8 @@ static unsigned int free_old_xmit_skbs(sreturn tot_sgs; } -static int xmit_skb (struct virtnet_info *vi,struct sk_buff *skb) +static int xmit_skb (struct virtnet_info *vi,struct sk_buff *skb, +struct virtqueue *svq,struct scatterlist *tx_sg) {struct skb_vnet_hdr *hdr =skb_vnet_hdr (skb);const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;@@ -548,12 +561,12 @@ static int xmit_skb(struct virtnet_info/* Encode metadata header at front. */ if (vi->mergeable_rx_bufs) -sg_set_buf (vi->tx_sg, &hdr->mhdr,sizeof hdr->mhdr); +sg_set_buf (tx_sg, &hdr->mhdr,sizeof hdr->mhdr);else -sg_set_buf (vi->tx_sg, &hdr->hdr,sizeof hdr->hdr); +sg_set_buf (tx_sg, &hdr->hdr,sizeof hdr->hdr); - hdr->num_sg =skb_to_sgvec (skb, vi->tx_sg +1 ,0 , skb->len) +1 ; -return virtqueue_add_buf (vi->svq, vi->tx_sg, hdr->num_sg, + hdr->num_sg =skb_to_sgvec (skb, tx_sg +1 ,0 , skb->len) +1 ; +return virtqueue_add_buf (svq, tx_sg, hdr->num_sg,0 , skb); }@@ -561,31 +574,34 @@ static netdev_tx_t start_xmit(struct sk_ {struct virtnet_info *vi =netdev_priv (dev);int capacity; +int qnum =skb_get_queue_mapping (skb); +struct virtqueue *svq = vi->sq[qnum].svq;/* Free up any pending old buffers before queueing new ones. */ -free_old_xmit_skbs (vi); +free_old_xmit_skbs (vi, svq);/* Try to transmit */ - capacity =xmit_skb (vi, skb); + capacity =xmit_skb (vi, skb, svq, vi->sq[qnum].tx_sg);/* This can happen with OOM and indirect buffers. */ if (unlikely (capacity <0 )) {if (net_ratelimit ()) {if (likely (capacity == -ENOMEM )) {dev_warn (&dev->dev, -"TX queue failure: out of memory\n" ); +"TXQ (%d) failure: out of memory\n" , + qnum); }else { dev->stats.tx_fifo_errors++;dev_warn (&dev->dev, -"Unexpected TX queue failure: %d\n" , - capacity); +"Unexpected TXQ (%d) failure: %d\n" , + qnum, capacity); } } dev->stats.tx_dropped++;kfree_skb (skb);return NETDEV_TX_OK ; } -virtqueue_kick (vi->svq); +virtqueue_kick (svq);/* Don't wait up for transmitted skbs to be freed. */ skb_orphan (skb);@@ -594,13 +610,13 @@ static netdev_tx_t start_xmit(struct sk_/* Apparently nice girls don't return TX_BUSY; stop the queue * before it gets out of hand. Naturally, this wastes entries. */ if (capacity <2 +MAX_SKB_FRAGS ) { -netif_stop_queue (dev); -if (unlikely (!virtqueue_enable_cb (vi->svq))) { +netif_stop_subqueue (dev, qnum); +if (unlikely (!virtqueue_enable_cb (svq))) {/* More just got used, free them then recheck. */ - capacity +=free_old_xmit_skbs (vi); + capacity +=free_old_xmit_skbs (vi, svq);if (capacity >=2 +MAX_SKB_FRAGS ) { -netif_start_queue (dev); -virtqueue_disable_cb (vi->svq); +netif_start_subqueue (dev, qnum); +virtqueue_disable_cb (svq); } } }@@ -871,10 +887,10 @@ static void virtnet_update_status(structif (vi->status &VIRTIO_NET_S_LINK_UP ) {netif_carrier_on (vi->dev); -netif_wake_queue (vi->dev); +netif_tx_wake_all_queues (vi->dev); }else {netif_carrier_off (vi->dev); -netif_stop_queue (vi->dev); +netif_tx_stop_all_queues (vi->dev); } }@@ -885,18 +901,112 @@ static void virtnet_config_changed(strucvirtnet_update_status (vi); }
drivers/net/virtio_net.c static void virtnet_config_changed(struc · line 913
Quoted 10 lines
drivers/net/virtio_net.c +10
-0+#define MAX_DEVICE_NAME 16 +static int initialize_vqs (struct virtnet_info *vi,int numtxqs) +{ +vq_callback_t **callbacks; +struct virtqueue **vqs; +int i, err = -ENOMEM ; +int totalvqs; +char **names; + +/* Allocate send queues */
no check on numtxqs ? Hmm...
Please then use kcalloc(numtxqs, sizeof(*vi->sq), GFP_KERNEL) so that some check is done for you ;)
Quoted 22 lines
drivers/net/virtio_net.c +22
-0+ vi->sq =kzalloc (numtxqs *sizeof (*vi->sq),GFP_KERNEL ); +if (!vi->sq) + goto out; + +/* setup initial send queue parameters */ +for (i =0 ; i < numtxqs; i++) +sg_init_table (vi->sq[i].tx_sg,ARRAY_SIZE (vi->sq[i].tx_sg)); + +/* +* We expect 1 RX virtqueue followed by 'numtxqs' TX virtqueues, and +* optionally one control virtqueue. +*/ + totalvqs =1 + numtxqs + +virtio_has_feature (vi->vdev,VIRTIO_NET_F_CTRL_VQ ); + +/* Setup parameters for find_vqs */ + vqs =kmalloc (totalvqs *sizeof (*vqs),GFP_KERNEL ); + callbacks =kmalloc (totalvqs *sizeof (*callbacks),GFP_KERNEL ); + names =kzalloc (totalvqs *sizeof (*names),GFP_KERNEL ); +if (!vqs || !callbacks || !names) + goto free_mem; +
Re: [v2 RFC PATCH 2/4] Changes for virtio-net
Eric Dumazet <[email protected]> wrote on 09/17/2010 03:55:54 PM:
Quoted 9 lines
drivers/net/virtio_net.c module_param(gso, bool, 0444); · line 49
Quoted 7 lines
drivers/net/virtio_net.c +7
-0+/* Our representation of a send virtqueue */ +struct send_queue { +struct virtqueue *svq; + +/* TX: fragments + linear part + virtio header */ +struct scatterlist tx_sg[MAX_SKB_FRAGS +2 ]; +};You probably want
____cacheline_aligned_in_smp
I had tried this and mentioned this in Patch 0/4:
"2. Cache-align data structures: I didn't see any BW/SD improvement
after making the sq's (and similarly for vhost) cache-aligned statically:
struct virtnet_info {
...
struct send_queue sq[16] ____cacheline_aligned_in_smp;
...
};"
I am not sure why this made no difference?
Quoted 10 lines
drivers/net/virtio_net.c module_param(gso, bool, 0444); · line 57
Quoted 9 lines
drivers/net/virtio_net.c +5
-1+struct virtnet_info {struct virtio_device *vdev; -struct virtqueue *rvq, *svq, *cvq; +int numtxqs;/* Number of tx queues */ +struct send_queue *sq; +struct virtqueue *rvq; +struct virtqueue *cvq;struct net_device *dev;
Quoted 4 lines
struct napi will probably be dirtied by RX processing
You should make sure it doesnt dirty cache line of above (read mostly) fields
I am changing the layout of napi wrt other pointers in this patch, though the to-be-submitted RX patch does that. Should I do something for this TX-only patch?
Quoted 13 lines
drivers/net/virtio_net.c static void virtnet_config_changed(struc · line 913
Quoted 10 lines
drivers/net/virtio_net.c +10
-0+#define MAX_DEVICE_NAME 16 +static int initialize_vqs (struct virtnet_info *vi,int numtxqs) +{ +vq_callback_t **callbacks; +struct virtqueue **vqs; +int i, err = -ENOMEM ; +int totalvqs; +char **names; + +/* Allocate send queues */ no check on numtxqs ? Hmm...
Quoted 2 lines
Please then use
kcalloc(numtxqs, sizeof(*vi->sq), GFP_KERNEL)so that some check is done for you ;)
Right! I need to re-introduce some limit. Rusty, should I simply add a check for a constant (like 256) here?
Thanks for your review, Eric!
- KK
Re: [v2 RFC PATCH 2/4] Changes for virtio-net
Quoted 13 lines
Krishna Kumar2/India/IBM@IBMIN
Sent by: [email protected]Quoted 10 lines
Quoted 9 lines
drivers/net/virtio_net.c +5
-1+struct virtnet_info {struct virtio_device *vdev; -struct virtqueue *rvq, *svq, *cvq; +int numtxqs;/* Number of tx queues */ +struct send_queue *sq; +struct virtqueue *rvq; +struct virtqueue *cvq;struct net_device *dev;
Quoted 8 lines
Quoted 4 lines
struct napi will probably be dirtied by RX processing
You should make sure it doesnt dirty cache line of above (read mostly) fields
I am changing the layout of napi wrt other pointers in this patch, though the to-be-submitted RX patch does that. Should I do something for this TX-only patch?
Sorry, I think my sentence is not clear! I will make this change (and also cache-line align the send queues), test and let you know the result.
Thanks,
- KK
RFC PATCH v2 3/4 Changes for vhost
In reply to Krishna KumarChanges for mq vhost.
vhost_net_open is changed to allocate a vhost_net and return. The remaining initializations are delayed till SET_OWNER. SET_OWNER is changed so that the argument is used to figure out how many txqs to use. Unmodified qemu's will pass NULL, so this is recognized and handled as numtxqs=1.
Besides changing handle_tx to use 'vq', this patch also changes handle_rx to take vq as parameter. The mq RX patch requires this change, but till then it is consistent (and less confusing) to make the interfaces for handling rx and tx similar.
--- drivers/vhost/net.c | 273 ++++++++++++++++++++++++++-------------- drivers/vhost/vhost.c | 186 +++++++++++++++++++-------- drivers/vhost/vhost.h | 16 +- 3 files changed, 324 insertions(+), 151 deletions(-)
drivers/vhost/vhost.h +11 -5
@@ -40,11 +40,11 @@ struct vhost_poll {wait_queue_t wait;struct vhost_work work;unsigned long mask; -struct vhost_dev *dev; +struct vhost_virtqueue *vq;/* points back to vq */ };void vhost_poll_init (struct vhost_poll *poll,vhost_work_fn_t fn, -unsigned long mask,struct vhost_dev *dev); +unsigned long mask,struct vhost_virtqueue *vq);void vhost_poll_start (struct vhost_poll *poll,struct file *file);void vhost_poll_stop (struct vhost_poll *poll);void vhost_poll_flush (struct vhost_poll *poll);
@@ -110,6 +110,10 @@ struct vhost_virtqueue {/* Log write descriptors */ void __user *log_base;struct vhost_log log[VHOST_NET_MAX_SG ]; +struct task_struct *worker;/* vhost for this vq, shared btwn RX/TX */ +spinlock_t *work_lock; +struct list_head *work_list; +int qnum;/* 0 for RX, 1 -> n-1 for TX */ };struct vhost_dev {
@@ -124,11 +128,13 @@ struct vhost_dev {int nvqs;struct file *log_file;struct eventfd_ctx *log_ctx; -spinlock_t work_lock; -struct list_head work_list; -struct task_struct *worker; +spinlock_t *work_lock; +struct list_head *work_list; }; +int vhost_setup_vqs (struct vhost_dev *dev,int numtxqs); +void vhost_free_vqs (struct vhost_dev *dev); +int vhost_get_num_threads (int nvqs);long vhost_dev_init (struct vhost_dev *,struct vhost_virtqueue *vqs,int nvqs);long vhost_dev_check_owner (struct vhost_dev *);long vhost_dev_reset_owner (struct vhost_dev *);
drivers/vhost/net.c +182 -91
449 lines not on this page · raw message
drivers/vhost/vhost.c +131 -55
316 lines not on this page · raw message
RFC PATCH v2 4/4 qemu changes
In reply to Krishna KumarChanges in qemu to support mq TX.
--- hw/vhost.c | 8 ++- hw/vhost.h | 2 hw/vhost_net.c | 16 +++++-- hw/vhost_net.h | 2 hw/virtio-net.c | 97 ++++++++++++++++++++++++++++++---------------- hw/virtio-net.h | 2 hw/virtio-pci.c | 2 net.c | 17 ++++++++ net.h | 1 net/tap.c | 27 ++++++++++-- 10 files changed, 129 insertions(+), 45 deletions(-)
hw/vhost.c +6 -2
@@ -599,23 +599,27 @@ static void vhost_virtqueue_cleanup(stru 0, virtio_queue_get_desc_size(vdev, idx)); } -int vhost_dev_init (struct vhost_dev *hdev,int devfd) +int vhost_dev_init (struct vhost_dev *hdev,int devfd,int numtxqs) {uint64_t features;int r;if (devfd >=0 ) { hdev->control = devfd; + hdev->nvqs =2 ; }else { hdev->control =open ("/dev/vhost-net" ,O_RDWR );if (hdev->control <0 ) {return -errno; } } - r =ioctl (hdev->control,VHOST_SET_OWNER ,NULL ); + + r =ioctl (hdev->control,VHOST_SET_OWNER , numtxqs);if (r <0 ) { goto fail; } + hdev->nvqs = numtxqs +1 ; + r =ioctl (hdev->control,VHOST_GET_FEATURES , &features);if (r <0 ) { goto fail;
hw/vhost.h +1 -1
@@ -40,7 +40,7 @@ struct vhost_dev {unsigned long long log_size; }; -int vhost_dev_init (struct vhost_dev *hdev,int devfd); +int vhost_dev_init (struct vhost_dev *hdev,int devfd,int nvqs);void vhost_dev_cleanup (struct vhost_dev *hdev);int vhost_dev_start (struct vhost_dev *hdev,VirtIODevice *vdev);void vhost_dev_stop (struct vhost_dev *hdev,VirtIODevice *vdev);
hw/vhost_net.c +11 -5
@@ -36,7 +36,8 @@struct vhost_net {struct vhost_dev dev; -struct vhost_virtqueue vqs[2 ]; +struct vhost_virtqueue *vqs; +int nvqs;int backend;VLANClientState *vc; };
@@ -76,7 +77,8 @@ static int vhost_net_get_fd(VLANClientSt } } -struct vhost_net *vhost_net_init (VLANClientState *backend,int devfd) +struct vhost_net *vhost_net_init (VLANClientState *backend,int devfd, +int numtxqs) {int r;struct vhost_net *net =qemu_malloc (sizeof *net);
@@ -93,10 +95,14 @@ struct vhost_net *vhost_net_init(VLANCli (1 <<VHOST_NET_F_VIRTIO_NET_HDR ); net->backend = r; - r =vhost_dev_init (&net->dev, devfd); + r =vhost_dev_init (&net->dev, devfd, numtxqs);if (r <0 ) { goto fail; } + + net->nvqs = numtxqs +1 ; + net->vqs =qemu_malloc (net->nvqs * (sizeof *net->vqs)); +if (~net->dev.features & net->dev.backend_features) {fprintf (stderr,"vhost lacks feature mask %" PRIu64" for backend\n" , (uint64_t )(~net->dev.features & net->dev.backend_features));
@@ -118,7 +124,6 @@ int vhost_net_start(struct vhost_net *nestruct vhost_vring_file file = { };int r; - net->dev.nvqs =2 ; net->dev.vqs = net->vqs; r =vhost_dev_start (&net->dev, dev);if (r <0 ) {
@@ -166,7 +171,8 @@ void vhost_net_cleanup(struct vhost_net qemu_free(net); }#else -struct vhost_net *vhost_net_init (VLANClientState *backend,int devfd) +struct vhost_net *vhost_net_init (VLANClientState *backend,int devfd, +int nvqs) {return NULL ; }
hw/vhost_net.h +1 -1
@@ -6,7 +6,7 @@struct vhost_net ;typedef struct vhost_net VHostNetState ; -VHostNetState *vhost_net_init (VLANClientState *backend,int devfd); +VHostNetState *vhost_net_init (VLANClientState *backend,int devfd,int nvqs);int vhost_net_start (VHostNetState *net,VirtIODevice *dev);void vhost_net_stop (VHostNetState *net,VirtIODevice *dev);
hw/virtio-net.c +65 -32
228 lines not on this page · raw message
hw/virtio-net.h +2 -0
@@ -44,6 +44,7 @@#define VIRTIO_NET_F_CTRL_RX 18/* Control channel RX mode support */ #define VIRTIO_NET_F_CTRL_VLAN 19/* Control channel VLAN filtering */ #define VIRTIO_NET_F_CTRL_RX_EXTRA 20/* Extra RX mode control support */ +#define VIRTIO_NET_F_NUMTXQS 21/* Supports multiple TX queues */ #define VIRTIO_NET_S_LINK_UP 1/* Link is up */
@@ -58,6 +59,7 @@ struct virtio_net_configuint8_t mac[ETH_ALEN ];/* See VIRTIO_NET_F_STATUS and VIRTIO_NET_S_* above */ uint16_t status; +uint16_t numtxqs;/* number of transmit queues */ } __attribute__((packed));/* This is the first element of the scatter-gather list. If you don't
hw/virtio-pci.c +2 -0
@@ -99,6 +99,7 @@ typedef struct {uint32_t addr;uint32_t class_code;uint32_t nvectors; +uint32_t mq;BlockConf block;NICConf nic;uint32_t host_features;
@@ -722,6 +723,7 @@ static PCIDeviceInfo virtio_info[] = { .romfile = "pxe-virtio.bin", .qdev.props = (Property []) {DEFINE_PROP_UINT32 ("vectors" , VirtIOPCIProxy, nvectors,3 ), +DEFINE_PROP_UINT32 ("mq" , VirtIOPCIProxy, mq,1 ),DEFINE_VIRTIO_NET_FEATURES (VirtIOPCIProxy, host_features),DEFINE_NIC_PROPERTIES (VirtIOPCIProxy, nic),DEFINE_PROP_END_OF_LIST (),
net/tap.c +23 -4
@@ -299,13 +299,14 @@ static NetClientInfo net_tap_info = {static TAPState *net_tap_fd_init (VLANState *vlan,const char *model,const char *name, -int fd, +int fd,int numtxqs,int vnet_hdr) {VLANClientState *nc;TAPState *s; nc =qemu_new_net_client (&net_tap_info, vlan,NULL , model, name); + nc->numtxqs = numtxqs; s =DO_UPCAST (TAPState, nc, nc);
@@ -403,6 +404,24 @@ int net_init_tap(QemuOpts *opts, Monitor {TAPState *s;int fd, vnet_hdr =0 ; +int vhost; +int numtxqs =1 ; + + vhost =qemu_opt_get_bool (opts,"vhost" ,0 ); + +/* +* We support multiple tx queues if: +* 1. smp > 1 +* 2. vhost=on +* 3. mq=on +* In this case, #txqueues = #cpus. This value can be changed by +* using the "numtxqs" option. +*/ +if (vhost && smp_cpus >1 ) { +if (qemu_opt_get_bool (opts,"mq" ,0 )) { + numtxqs =qemu_opt_get_number (opts,"numtxqs" , smp_cpus); + } + }if (qemu_opt_get (opts,"fd" )) { if (qemu_opt_get(opts, "ifname") ||
@@ -436,7 +455,7 @@ int net_init_tap(QemuOpts *opts, Monitor } } - s =net_tap_fd_init (vlan,"tap" , name, fd, vnet_hdr); + s =net_tap_fd_init (vlan,"tap" , name, fd, numtxqs, vnet_hdr);if (!s) {close (fd);return -1 ;
@@ -465,7 +484,7 @@ int net_init_tap(QemuOpts *opts, Monitor } } -if (qemu_opt_get_bool (opts,"vhost" , !!qemu_opt_get (opts,"vhostfd" ))) { +if (vhost) {int vhostfd, r;if (qemu_opt_get (opts,"vhostfd" )) { r =net_handle_fd_param (mon,qemu_opt_get (opts,"vhostfd" ));
@@ -476,7 +495,7 @@ int net_init_tap(QemuOpts *opts, Monitor }else { vhostfd =-1 ; } - s->vhost_net =vhost_net_init (&s->nc, vhostfd); + s->vhost_net =vhost_net_init (&s->nc, vhostfd, numtxqs);if (!s->vhost_net) {error_report ("vhost-net requested but could not be initialized" );return -1 ;
net.c +17 -0
@@ -814,6 +814,15 @@ static int net_init_nic(QemuOpts *opts,return -1 ; } +if (nd->netdev->numtxqs >1 && nd->nvectors ==DEV_NVECTORS_UNSPECIFIED ) { +/* +* User specified mq for guest, but no "vectors=", tune +* it automatically to 'numtxqs' TX + 1 RX + 1 controlq. +*/ + nd->nvectors = nd->netdev->numtxqs +1 +1 ; +monitor_printf (mon,"nvectors tuned to %d\n" , nd->nvectors); + } + nd->used =1 ; nb_nics++;
@@ -957,6 +966,14 @@ static const struct { }, #ifndef _WIN32 { + .name = "mq", + .type = QEMU_OPT_BOOL, + .help = "enable multiqueue on network i/f", + }, { + .name = "numtxqs", + .type = QEMU_OPT_NUMBER, + .help = "optional number of TX queues, if mq is enabled", + }, { .name = "fd", .type = QEMU_OPT_STRING, .help = "file descriptor of an already opened tap",
net.h +1 -0
@@ -62,6 +62,7 @@ struct VLANClientState {struct VLANState *vlan;VLANClientState *peer;NetQueue *send_queue; +int numtxqs;char *model;char *name;char info_str[256 ];
Quoted 35 lines · Krishna Kumar
Following patches implement transmit MQ in virtio-net. Also included is the user qemu changes. MQ is disabled by default unless qemu specifies it.
1. This feature was first implemented with a single vhost. Testing showed 3-8% performance gain for upto 8 netperf sessions (and sometimes 16), but BW dropped with more sessions. However, adding more vhosts improved BW significantly all the way to 128 sessions. Multiple vhost is implemented in-kernel by passing an argument to
SET_OWNER(retaining backward compatibility). The vhost patch adds 173 source lines (incl comments).
2. BW -> CPU/SD equation: Average TCP performance increased 23% compared to almost 70% for earlier patch (with unrestricted #vhosts). SD improved -4.2% while it had increased 55% for the earlier patch. Increasing #vhosts has it's pros and cons, but this patch lays emphasis on reducing CPU utilization. Another option could be a tunable to select number of vhosts threads.
3. Interoperability: Many combinations, but not all, of qemu, host, guest tested together. Tested with multiple i/f's on guest, with both mq=on/off, vhost=on/off, etc.Changes from rev1:
------------------
1. Movequeue_indexfromvirtio_pci_vq_infoto virtqueue, and resulting changes to existing code and to the patch.
2. virtio-net probe usesvirtio_config_val.
3. Remove constants:VIRTIO_MAX_TXQS,MAX_VQS, all arrays allocated on stack, etc.
4. Restrict number of vhost threads to 2 - I get much better cpu/sd results (without any tuning) with low number of vhost threads. Higher vhosts gives better average BW performance (from average of 45%), but SD increases significantly (90%).
5. Working of vhost threads changes, eg for numtxqs=4:
Quoted 5 lines
vhost-0: handles RX
vhost-1: handles TX[0]
vhost-0: handles TX[1]
vhost-1: handles TX[2]
vhost-0: handles TX[3]This doesn't look symmetrical.
TCP flows that go via TX(1,3) use the same vhost thread for RX packets, whereas flows via TX(0,2) use a different vhost thread.
Thanks
Sridhar
Michael S. Tsirkin
maintainerQuoted 2 lines · Krishna Kumar
For 1 TCP netperf, I ran 7 iterations and summed it. Explanation for degradation for 1 stream case:
Could you document how exactly do you measure multistream bandwidth:
netperf flags, etc?
Quoted 1 line
1. Without any tuning, BW falls -6.5%.
Any idea where does this come from?
Do you see more TX interrupts? RX interrupts? Exits? Do interrupts bounce more between guest CPUs?
Quoted 13 lines
2. When vhosts on server were bound to CPU0, BW was as good as with original code.
3. When new code was started with numtxqs=1 (or mq=off, which is the default), there was no degradation.Next steps:
-----------
1. MQ RX patch is also complete - plan to submit once TX is OK (as well as after identifying bandwidth degradations for some test cases).
2. Cache-align data structures: I didn't see any BW/SD improvement after making the sq's (and similarly for vhost) cache-aligned statically:
Quoted 6 lines
struct virtnet_info { ... struct send_queue sq[16] ____cacheline_aligned_in_smp; ... };3. Migration is not tested.
4. Identify reasons for single netperf BW regression.
5. Test perf in more scenarious:
small packets
host -> guest
guest <-> external
in last case:
find some other way to measure host CPU utilization, try multiqueue and single queue devices
6. Use above to figure out what is a sane default for numtxqs.
Quoted 5 lines
Review/feedback appreciated.
Signed-off-by Krishna Kumar <[email protected]>---
"Michael S. Tsirkin" <[email protected]> wrote on 09/19/2010 06:14:43 PM:
Quoted 2 lines
Could you document how exactly do you measure multistream bandwidth:
netperf flags, etc?
All results were without any netperf flags or system tuning:
for i in $list
do
netperf -c -C -l 60 -H 192.168.122.1 > /tmp/netperf.$$.$i &
done
wait
Another script processes the result files. It also displays the start time/end time of each iteration to make sure skew due to parallel netperfs is minimal.
I changed the vhost functionality once more to try to get the best model, the new model being:
1. #numtxqs=1 -> #vhosts=1, this thread handles both RX/TX.
2. #numtxqs>1 -> vhost[0] handles RX and vhost[1-MAX] handles TX[0-n], where MAX is 4. Beyond numtxqs=4, the remaining TX queues are handled by vhost threads in round-robin fashion.
Results from here on are with these changes, and only "tuning" is to set each vhost's affinity to CPUs[0-3] ("taskset -p f <vhost-pids>").
Quoted 4 lines
Any idea where does this come from?
Do you see more TX interrupts? RX interrupts? Exits? Do interrupts bounce more between guest CPUs?
4. Identify reasons for single netperf BW regression.
After testing various combinations of #txqs, #vhosts, #netperf sessions, I think the drop for 1 stream is due to TX and RX for a flow being processed on different cpus. I did two more tests:
1. Pin vhosts to same CPU:
- BW drop is much lower for 1 stream case (- 5 to -8% range)
- But performance is not so high for more sessions.
2. Changed vhost to be single threaded:
- No degradation for 1 session, and improvement for upto 8, sometimes 16 streams (5-12%).
- BW degrades after that, all the way till 128 netperf sessions.
- But overall CPU utilization improves.
Summary of the entire run (for 1-128 sessions):
txq=4: BW: (-2.3) CPU: (-16.5) RCPU: (-5.3)
txq=16: BW: (-1.9) CPU: (-24.9) RCPU: (-9.6)
I don't see any reasons mentioned above. However, for higher number of netperf sessions, I see a big increase in retransmissions:
_______________________________________
#netperf ORG NEW
BW (#retr) BW (#retr)
_______________________________________
1 70244 (0) 64102 (0) 4 21421 (0) 36570 (416) 8 21746 (0) 38604 (148) 16 21783 (0) 40632 (464) 32 22677 (0) 37163 (1053) 64 23648 (4) 36449 (2197) 128 23251 (2) 31676 (3185)
_______________________________________
Single netperf case didn't have any retransmissions so that is not the cause for drop. I tested ixgbe (MQ):
___________________________________________________________
#netperf ixgbe ixgbe (pin intrs to cpu#0 on
both server/client)
BW (#retr) BW (#retr)
___________________________________________________________
1 3567 (117) 6000 (251) 2 4406 (477) 6298 (725) 4 6119 (1085) 7208 (3387) 8 6595 (4276) 7381 (15296) 16 6651 (11651) 6856 (30394)
___________________________________________________________
Quoted 2 lines
5. Test perf in more scenarious:
small packets
512 byte packets - BW drop for upto 8 (sometimes 16) netperf sessions, but increases with #sessions:
_______________________________________________________________________________
# BW1 BW2 (%) CPU1 CPU2 (%) RCPU1 RCPU2 (%)
_______________________________________________________________________________
1 4043 3800 (-6.0) 50 50 (0) 86 98 (13.9) 2 8358 7485 (-10.4) 153 178 (16.3) 230 264 (14.7) 4 20664 13567 (-34.3) 448 490 (9.3) 530 624 (17.7) 8 25198 17590 (-30.1) 967 1021 (5.5) 1085 1257 (15.8) 16 23791 24057 (1.1) 1904 2220 (16.5) 2156 2578 (19.5) 24 23055 26378 (14.4) 2807 3378 (20.3) 3225 3901 (20.9) 32 22873 27116 (18.5) 3748 4525 (20.7) 4307 5239 (21.6) 40 22876 29106 (27.2) 4705 5717 (21.5) 5388 6591 (22.3) 48 23099 31352 (35.7) 5642 6986 (23.8) 6475 8085 (24.8) 64 22645 30563 (34.9) 7527 9027 (19.9) 8619 10656 (23.6) 80 22497 31922 (41.8) 9375 11390 (21.4) 10736 13485 (25.6) 96 22509 32718 (45.3) 11271 13710 (21.6) 12927 16269 (25.8) 128 22255 32397 (45.5) 15036 18093 (20.3) 17144 21608 (26.0)
_______________________________________________________________________________
SUM: BW: (16.7) CPU: (20.6) RCPU: (24.3)
_______________________________________________________________________________
Quoted 1 line
host -> guest
_______________________________________________________________________________
# BW1 BW2 (%) CPU1 CPU2 (%) RCPU1 RCPU2 (%)
_______________________________________________________________________________
*1 70706 90398 (27.8) 300 327 (9.0) 140 175 (25.0) 2 20951 21937 (4.7) 188 196 (4.2) 93 103 (10.7) 4 19952 25281 (26.7) 397 496 (24.9) 210 304 (44.7) 8 18559 24992 (34.6) 802 1010 (25.9) 439 659 (50.1) 16 18882 25608 (35.6) 1642 2082 (26.7) 953 1454 (52.5) 24 19012 26955 (41.7) 2465 3153 (27.9) 1452 2254 (55.2) 32 19846 26894 (35.5) 3278 4238 (29.2) 1914 3081 (60.9) 40 19704 27034 (37.2) 4104 5303 (29.2) 2409 3866 (60.4) 48 19721 26832 (36.0) 4924 6418 (30.3) 2898 4701 (62.2) 64 19650 26849 (36.6) 6595 8611 (30.5) 3975 6433 (61.8) 80 19432 26823 (38.0) 8244 10817 (31.2) 4985 8165 (63.7) 96 20347 27886 (37.0) 9913 13017 (31.3) 5982 9860 (64.8) 128 19108 27715 (45.0) 13254 17546 (32.3) 8153 13589 (66.6)
_______________________________________________________________________________
SUM: BW: (32.4) CPU: (30.4) RCPU: (62.6)
_______________________________________________________________________________
*: Sum over 7 iterations, remaining test cases are sum over 2 iterations
Quoted 1 line
guest <-> external
I haven't done this right now since I don't have a setup. I guess it would be limited by wire speed and gains may not be there. I will try to do this later when I get the setup.
Quoted 4 lines
in last case:
find some other way to measure host CPU utilization, try multiqueue and single queue devices
6. Use above to figure out what is a sane default for numtxqs
A. Summary for default I/O (16K):
#txqs=2 (#vhost=3): BW: (37.6) CPU: (69.2) RCPU: (40.8) #txqs=4 (#vhost=5): BW: (36.9) CPU: (60.9) RCPU: (25.2) #txqs=8 (#vhost=5): BW: (41.8) CPU: (50.0) RCPU: (15.2) #txqs=16 (#vhost=5): BW: (40.4) CPU: (49.9) RCPU: (10.0)
B. Summary for 512 byte I/O:
#txqs=2 (#vhost=3): BW: (31.6) CPU: (35.7) RCPU: (28.6) #txqs=4 (#vhost=5): BW: (5.7) CPU: (27.2) RCPU: (22.7) #txqs=8 (#vhost=5): BW: (-.6) CPU: (25.1) RCPU: (22.5) #txqs=16 (#vhost=5): BW: (-6.6) CPU: (24.7) RCPU: (21.7)
Summary:
1. Average BW increase for regular I/O is best for #txq=16 with the least CPU utilization increase.
2. The average BW for 512 byte I/O is best for lower #txq=2. For higher #txqs, BW increased only after a particular #netperf sessions - in my testing that limit was 32 netperf sessions.
3. Multiple txq for guest by itself doesn't seem to have any issues. Guest CPU% increase is slightly higher than BW improvement. I think it is true for all mq drivers since more paths run in parallel upto the device instead of sleeping and allowing one thread to send all packets via qdisc_restart.
4. Having high number of txqs gives better gains and reduces cpu util on the guest and the host.
5. MQ is intended for server loads. MQ should probably not be explicitly specified for client systems.
6. No regression with numtxqs=1 (or if mq option is not used) in any testing scenario.
I will send the v3 patch within a day after some more testing.
Thanks,
- KK
Michael S. Tsirkin
maintainerQuoted 25 lines · On Tue, Oct 05, 2010 at 04:10:00PM +0530, Krishna Kumar2 wrote:
"Michael S. Tsirkin" <[email protected]> wrote on 09/19/2010 06:14:43 PM:
Quoted 2 lines
Could you document how exactly do you measure multistream bandwidth:
netperf flags, etc?All results were without any netperf flags or system tuning:
for i in $list
do
netperf -c -C -l 60 -H 192.168.122.1 >/tmp/netperf.$$.$i&
done
wait
Another script processes the result files. It also displays the start time/end time of each iteration to make sure skew due to parallel netperfs is minimal.I changed the vhost functionality once more to try to get the best model, the new model being:
1. #numtxqs=1 -> #vhosts=1, this thread handles both RX/TX.
2. #numtxqs>1 -> vhost[0] handles RX and vhost[1-MAX] handles TX[0-n], where MAX is 4. Beyond numtxqs=4, the remaining TX queues are handled by vhost threads in round-robin fashion.Results from here on are with these changes, and only "tuning" is to set each vhost's affinity to CPUs[0-3] ("taskset -p f <vhost-pids>").
Quoted 8 lines
Quoted 4 lines
Any idea where does this come from?
Do you see more TX interrupts? RX interrupts? Exits? Do interrupts bounce more between guest CPUs?
4. Identify reasons for single netperf BW regression.After testing various combinations of #txqs, #vhosts, #netperf sessions, I think the drop for 1 stream is due to TX and RX for a flow being processed on different cpus.
Right. Can we fix it?
Quoted 13 lines
I did two more tests:
1. Pin vhosts to same CPU:
- BW drop is much lower for 1 stream case (- 5 to -8% range)
- But performance is not so high for more sessions.
2. Changed vhost to be single threaded:
- No degradation for 1 session, and improvement for upto 8, sometimes 16 streams (5-12%).
- BW degrades after that, all the way till 128 netperf sessions.
- But overall CPU utilization improves.
Summary of the entire run (for 1-128 sessions):
txq=4: BW: (-2.3) CPU: (-16.5) RCPU: (-5.3)
txq=16: BW: (-1.9) CPU: (-24.9) RCPU: (-9.6)
Quoted 2 lines
I don't see any reasons mentioned above. However, for higher number of netperf sessions, I see a big increase in retransmissions:
Hmm, ok, and do you see any errors?
Quoted 20 lines
_______________________________________
#netperf ORG NEW
BW (#retr) BW (#retr)
_______________________________________1 70244 (0) 64102 (0) 4 21421 (0) 36570 (416) 8 21746 (0) 38604 (148) 16 21783 (0) 40632 (464) 32 22677 (0) 37163 (1053) 64 23648 (4) 36449 (2197) 128 23251 (2) 31676 (3185)_______________________________________
Single netperf case didn't have any retransmissions so that is not the cause for drop. I tested ixgbe (MQ):
___________________________________________________________
#netperf ixgbe ixgbe (pin intrs to cpu#0 on
both server/client)
BW (#retr) BW (#retr)
___________________________________________________________
Quoted 5 lines
1 3567 (117) 6000 (251) 2 4406 (477) 6298 (725) 4 6119 (1085) 7208 (3387) 8 6595 (4276) 7381 (15296) 16 6651 (11651) 6856 (30394)
Interesting.
You are saying we get much more retransmissions with physical nic as well?
Quoted 49 lines
___________________________________________________________
Quoted 2 lines
5. Test perf in more scenarious:
small packets512 byte packets - BW drop for upto 8 (sometimes 16) netperf sessions, but increases with #sessions:
_______________________________________________________________________________
# BW1 BW2 (%) CPU1 CPU2 (%) RCPU1 RCPU2 (%)
_______________________________________________________________________________1 4043 3800 (-6.0) 50 50 (0) 86 98 (13.9) 2 8358 7485 (-10.4) 153 178 (16.3) 230 264 (14.7) 4 20664 13567 (-34.3) 448 490 (9.3) 530 624 (17.7) 8 25198 17590 (-30.1) 967 1021 (5.5) 1085 1257 (15.8) 16 23791 24057 (1.1) 1904 2220 (16.5) 2156 2578 (19.5) 24 23055 26378 (14.4) 2807 3378 (20.3) 3225 3901 (20.9) 32 22873 27116 (18.5) 3748 4525 (20.7) 4307 5239 (21.6) 40 22876 29106 (27.2) 4705 5717 (21.5) 5388 6591 (22.3) 48 23099 31352 (35.7) 5642 6986 (23.8) 6475 8085 (24.8) 64 22645 30563 (34.9) 7527 9027 (19.9) 8619 10656 (23.6) 80 22497 31922 (41.8) 9375 11390 (21.4) 10736 13485 (25.6) 96 22509 32718 (45.3) 11271 13710 (21.6) 12927 16269 (25.8) 128 22255 32397 (45.5) 15036 18093 (20.3) 17144 21608 (26.0)_______________________________________________________________________________
SUM: BW: (16.7) CPU: (20.6) RCPU: (24.3)
_______________________________________________________________________________Quoted 1 line
host -> guest
_______________________________________________________________________________
# BW1 BW2 (%) CPU1 CPU2 (%) RCPU1 RCPU2 (%)
_______________________________________________________________________________*1 70706 90398 (27.8) 300 327 (9.0) 140 175 (25.0) 2 20951 21937 (4.7) 188 196 (4.2) 93 103 (10.7) 4 19952 25281 (26.7) 397 496 (24.9) 210 304 (44.7) 8 18559 24992 (34.6) 802 1010 (25.9) 439 659 (50.1) 16 18882 25608 (35.6) 1642 2082 (26.7) 953 1454 (52.5) 24 19012 26955 (41.7) 2465 3153 (27.9) 1452 2254 (55.2) 32 19846 26894 (35.5) 3278 4238 (29.2) 1914 3081 (60.9) 40 19704 27034 (37.2) 4104 5303 (29.2) 2409 3866 (60.4) 48 19721 26832 (36.0) 4924 6418 (30.3) 2898 4701 (62.2) 64 19650 26849 (36.6) 6595 8611 (30.5) 3975 6433 (61.8) 80 19432 26823 (38.0) 8244 10817 (31.2) 4985 8165 (63.7) 96 20347 27886 (37.0) 9913 13017 (31.3) 5982 9860 (64.8) 128 19108 27715 (45.0) 13254 17546 (32.3) 8153 13589 (66.6)_______________________________________________________________________________
SUM: BW: (32.4) CPU: (30.4) RCPU: (62.6)
_______________________________________________________________________________
*: Sum over 7 iterations, remaining test cases are sum over 2 iterations
Quoted 5 lines
Quoted 1 line
guest <-> external
I haven't done this right now since I don't have a setup. I guess it would be limited by wire speed and gains may not be there. I will try to do this later when I get the setup.
OK but at least need to check that it does not hurt things.
Quoted 19 lines
Quoted 4 lines
in last case:
find some other way to measure host CPU utilization, try multiqueue and single queue devices
6. Use above to figure out what is a sane default for numtxqsA. Summary for default I/O (16K):
#txqs=2 (#vhost=3): BW: (37.6) CPU: (69.2) RCPU: (40.8) #txqs=4 (#vhost=5): BW: (36.9) CPU: (60.9) RCPU: (25.2) #txqs=8 (#vhost=5): BW: (41.8) CPU: (50.0) RCPU: (15.2) #txqs=16 (#vhost=5): BW: (40.4) CPU: (49.9) RCPU: (10.0)B. Summary for 512 byte I/O:
#txqs=2 (#vhost=3): BW: (31.6) CPU: (35.7) RCPU: (28.6) #txqs=4 (#vhost=5): BW: (5.7) CPU: (27.2) RCPU: (22.7) #txqs=8 (#vhost=5): BW: (-.6) CPU: (25.1) RCPU: (22.5) #txqs=16 (#vhost=5): BW: (-6.6) CPU: (24.7) RCPU: (21.7)Summary:
Quoted 16 lines
1. Average BW increase for regular I/O is best for #txq=16 with the least CPU utilization increase.
2. The average BW for 512 byte I/O is best for lower #txq=2. For higher #txqs, BW increased only after a particular #netperf sessions - in my testing that limit was 32 netperf sessions.
3. Multiple txq for guest by itself doesn't seem to have any issues. Guest CPU% increase is slightly higher than BW improvement. I think it is true for all mq drivers since more paths run in parallel upto the device instead of sleeping and allowing one thread to send all packets viaqdisc_restart.
4. Having high number of txqs gives better gains and reduces cpu util on the guest and the host.
5. MQ is intended for server loads. MQ should probably not be explicitly specified for client systems.
6. No regression with numtxqs=1 (or if mq option is not used) in any testing scenario.
Of course txq=1 can be considered a kind of fix, but if we know the issue is TX/RX flows getting bounced between CPUs, can we fix this? Workload-specific optimizations can only get us this far.
Quoted 6 lines
I will send the v3 patch within a day after some more testing.
Thanks,
- KK
"Michael S. Tsirkin" <[email protected]> wrote on 10/05/2010 11:53:23 PM:
Quoted 10 lines
Quoted 8 lines
Quoted 4 lines
Any idea where does this come from?
Do you see more TX interrupts? RX interrupts? Exits? Do interrupts bounce more between guest CPUs?
4. Identify reasons for single netperf BW regression.After testing various combinations of #txqs, #vhosts, #netperf sessions, I think the drop for 1 stream is due to TX and RX for a flow being processed on different cpus.
Right. Can we fix it?
I am not sure how to. My initial patch had one thread but gave small gains and ran into limitations once number of sessions became large.
Quoted 8 lines
Quoted 8 lines
I did two more tests:
1. Pin vhosts to same CPU:
- BW drop is much lower for 1 stream case (- 5 to -8% range)
- But performance is not so high for more sessions.
2. Changed vhost to be single threaded:
- No degradation for 1 session, and improvement for upto
8, sometimes 16 streams (5-12%).
- BW degrades after that, all the way till 128 netperf
sessions.
Quoted 9 lines
Quoted 7 lines
- But overall CPU utilization improves. Summary of the entire run (for 1-128 sessions):
txq=4: BW: (-2.3) CPU: (-16.5) RCPU: (-5.3)
txq=16: BW: (-1.9) CPU: (-24.9) RCPU: (-9.6)I don't see any reasons mentioned above. However, for higher number of netperf sessions, I see a big increase in retransmissions:
Hmm, ok, and do you see any errors?
I haven't seen any in any statistics, messages, etc. Also no retranmissions for txq=1.
Quoted 13 lines
Quoted 7 lines
Single netperf case didn't have any retransmissions so that is not the cause for drop. I tested ixgbe (MQ):
___________________________________________________________
#netperf ixgbe ixgbe (pin intrs to cpu#0 on
both server/client)
BW (#retr) BW (#retr)
___________________________________________________________Quoted 5 lines
1 3567 (117) 6000 (251) 2 4406 (477) 6298 (725) 4 6119 (1085) 7208 (3387) 8 6595 (4276) 7381 (15296) 16 6651 (11651) 6856 (30394)
Quoted 3 lines
Interesting.
You are saying we get much more retransmissions with physical nic as well?
Yes, with ixgbe. I re-ran with 16 netperfs running for 15 secs on both ixgbe and cxgb3 just now to reconfirm:
ixgbe: BW: 6186.85 SD/Remote: 135.711, 339.376 CPU/Remote: 79.99, 200.00,
Retrans: 545
cxgb3: BW: 8051.07 SD/Remote: 144.416, 260.487 CPU/Remote: 110.88, 200.00, Retrans: 0
However 64 netperfs for 30 secs gave:
ixgbe: BW: 6691.12 SD/Remote: 8046.617, 5259.992 CPU/Remote: 1223.86, 799.97, Retrans: 1424
cxgb3: BW: 7799.16 SD/Remote: 2589.875, 4317.013 CPU/Remote: 480.39 800.64, Retrans: 649
# ethtool -i eth4 driver: ixgbe version: 2.0.84-k2 firmware-version: 0.9-3 bus-info: 0000:1f:00.1
# ifconfig output:
RX packets:783241 errors:0 dropped:0 overruns:0 frame:0
TX packets:689533 errors:0 dropped:0 overruns:0 carrier:0
collisions:0 txqueuelen:1000# lspci output:
1f:00.0 Ethernet controller: Intel Corporation 82599EB 10-Gigabit Network
Connec
tion (rev 01)
Subsystem: Intel Corporation Ethernet Server Adapter X520-2
Flags: bus master, fast devsel, latency 0, IRQ 30
Memory at 98900000 (64-bit, prefetchable) [size=512K]
I/O ports at 2020 [size=32]
Memory at 98a00000 (64-bit, prefetchable) [size=16K]
Capabilities: [40] Power Management version 3
Capabilities: [50] MSI: Enable- Count=1/1 Maskable+ 64bit+
Capabilities: [70] MSI-X: Enable+ Count=64 Masked-
Capabilities: [a0] Express Endpoint, MSI 00
Capabilities: [100] Advanced Error Reporting
Capabilities: [140] Device Serial Number 00-1b-21-ff-ff-40-4a-b4
Capabilities: [150] Alternative Routing-ID Interpretation (ARI)
Capabilities: [160] Single Root I/O Virtualization (SR-IOV)
Kernel driver in use: ixgbe
Kernel modules: ixgbeQuoted 5 lines
Quoted 3 lines
I haven't done this right now since I don't have a setup. I guess it would be limited by wire speed and gains may not be there. I will try to do this later when I get the setup.
OK but at least need to check that it does not hurt things.
Yes, sure.
Quoted 15 lines
Quoted 15 lines
Summary:
1. Average BW increase for regular I/O is best for #txq=16 with the least CPU utilization increase.
2. The average BW for 512 byte I/O is best for lower #txq=2. For higher #txqs, BW increased only after a particular #netperf sessions - in my testing that limit was 32 netperf sessions.
3. Multiple txq for guest by itself doesn't seem to have any issues. Guest CPU% increase is slightly higher than BW improvement. I think it is true for all mq drivers since more paths run in parallel upto the device instead of sleeping and allowing one thread to send all packets viaqdisc_restart.
4. Having high number of txqs gives better gains and reduces cpu util on the guest and the host.
5. MQ is intended for server loads. MQ should probably not be
explicitly
Quoted 7 lines
Quoted 3 lines
specified for client systems.
6. No regression with numtxqs=1 (or if mq option is not used) in any testing scenario.Of course txq=1 can be considered a kind of fix, but if we know the issue is TX/RX flows getting bounced between CPUs, can we fix this? Workload-specific optimizations can only get us this far.
I will test with your patch tomorrow night once I am back.
Thanks,
- KK
Michael S. Tsirkin
maintainerQuoted 11 lines · On Wed, Oct 06, 2010 at 11:13:31PM +0530, Krishna Kumar2 wrote:
"Michael S. Tsirkin" <[email protected]> wrote on 10/05/2010 11:53:23 PM:
Quoted 9 lines
Quoted 8 lines
Quoted 4 lines
Any idea where does this come from?
Do you see more TX interrupts? RX interrupts? Exits? Do interrupts bounce more between guest CPUs?
4. Identify reasons for single netperf BW regression.After testing various combinations of #txqs, #vhosts, #netperf sessions, I think the drop for 1 stream is due to TX and RX for a flow being processed on different cpus.
Quoted 5 lines
Quoted 1 line
Right. Can we fix it?
I am not sure how to. My initial patch had one thread but gave small gains and ran into limitations once number of sessions became large.
Sure. We will need multiple RX queues, and have a single thread handle a TX and RX pair. Then we need to make sure packets from a given flow on TX land on the same thread on RX. As flows can be hashed differently, for this to work we'll have to expose this info in host/guest interface.
But since multiqueue implies host/guest ABI changes anyway, this point is moot.
BTW, an interesting approach could be using bonding and multiple virtio-net interfaces.
What are the disadvantages of such a setup? One advantage is it can be made to work in existing guests.
Quoted 14 lines
Quoted 8 lines
Quoted 8 lines
I did two more tests:
1. Pin vhosts to same CPU:
- BW drop is much lower for 1 stream case (- 5 to -8% range)
- But performance is not so high for more sessions.
2. Changed vhost to be single threaded:
- No degradation for 1 session, and improvement for upto
8, sometimes 16 streams (5-12%).
- BW degrades after that, all the way till 128 netperfsessions.
Quoted 5 lines
Quoted 5 lines
- But overall CPU utilization improves. Summary of the entire run (for 1-128 sessions):
txq=4: BW: (-2.3) CPU: (-16.5) RCPU: (-5.3)
txq=16: BW: (-1.9) CPU: (-24.9) RCPU: (-9.6)
Quoted 6 lines
Quoted 4 lines
Quoted 2 lines
I don't see any reasons mentioned above. However, for higher number of netperf sessions, I see a big increase in retransmissions:
Hmm, ok, and do you see any errors?
I haven't seen any in any statistics, messages, etc.
Herbert, could you help out debugging this increase in retransmissions please? Older mail on netdev in this thread has some numbers that seem to imply that we start hitting retransmissions much more as # of flows goes up.
Quoted 2 lines
Also no
retranmissions for txq=1.
While it's nice that we have this parameter, the need to choose between single stream and multi stream performance when you start the vm makes this patch much less interesting IMHO.
Signature · 1 line
MST
Quoted 21 lines · On Tuesday 05 October 2010, Krishna Kumar2 wrote:
After testing various combinations of #txqs, #vhosts, #netperf sessions, I think the drop for 1 stream is due to TX and RX for a flow being processed on different cpus. I did two more tests:
1. Pin vhosts to same CPU:
- BW drop is much lower for 1 stream case (- 5 to -8% range)
- But performance is not so high for more sessions.
2. Changed vhost to be single threaded:
- No degradation for 1 session, and improvement for upto 8, sometimes 16 streams (5-12%).
- BW degrades after that, all the way till 128 netperf sessions.
- But overall CPU utilization improves.
Summary of the entire run (for 1-128 sessions):
txq=4: BW: (-2.3) CPU: (-16.5) RCPU: (-5.3)
txq=16: BW: (-1.9) CPU: (-24.9) RCPU: (-9.6)I don't see any reasons mentioned above. However, for higher number of netperf sessions, I see a big increase in retransmissions:
_______________________________________
#netperf ORG NEW
BW (#retr) BW (#retr)
_______________________________________
Quoted 8 lines
1 70244 (0) 64102 (0) 4 21421 (0) 36570 (416) 8 21746 (0) 38604 (148) 16 21783 (0) 40632 (464) 32 22677 (0) 37163 (1053) 64 23648 (4) 36449 (2197) 128 23251 (2) 31676 (3185)_______________________________________
This smells like it could be related to a problem that Ben Greear found recently (see "macvlan: Enable qdisc backoff logic"). When the hardware is busy, used to just drop the packet. With Ben's patch, we return -EAGAIN to qemu (or vhost-net) to trigger a resend.
I suppose what we really should do is feed that condition back to the guest network stack and implement the backoff in there.
Arnd
Arnd Bergmann <[email protected]> wrote on 10/06/2010 05:49:00 PM:
Quoted 16 lines
Quoted 6 lines
I don't see any reasons mentioned above. However, for higher number of netperf sessions, I see a big increase in retransmissions:
_______________________________________
#netperf ORG NEW
BW (#retr) BW (#retr)
_______________________________________Quoted 8 lines
1 70244 (0) 64102 (0) 4 21421 (0) 36570 (416) 8 21746 (0) 38604 (148) 16 21783 (0) 40632 (464) 32 22677 (0) 37163 (1053) 64 23648 (4) 36449 (2197) 128 23251 (2) 31676 (3185)_______________________________________
Quoted 3 lines
This smells like it could be related to a problem that Ben Greear found recently (see "macvlan: Enable qdisc backoff logic"). When the hardware is busy, used to just drop the packet. With Ben's patch, we return
-EAGAIN
Quoted 4 lines
to qemu (or vhost-net) to trigger a resend.
I suppose what we really should do is feed that condition back to the guest network stack and implement the backoff in there.
Thanks for the pointer. I will take a look at this as I hadn't seen this patch earlier. Is there any way to figure out if this is the issue?
Thanks,
- KK
Quoted 22 lines · On Wednesday 06 October 2010 19:14:42 Krishna Kumar2 wrote:
Arnd Bergmann <[email protected]> wrote on 10/06/2010 05:49:00 PM:
Quoted 16 lines
Quoted 6 lines
I don't see any reasons mentioned above. However, for higher number of netperf sessions, I see a big increase in retransmissions:
_______________________________________
#netperf ORG NEW
BW (#retr) BW (#retr)
_______________________________________Quoted 8 lines
1 70244 (0) 64102 (0) 4 21421 (0) 36570 (416) 8 21746 (0) 38604 (148) 16 21783 (0) 40632 (464) 32 22677 (0) 37163 (1053) 64 23648 (4) 36449 (2197) 128 23251 (2) 31676 (3185)_______________________________________
Quoted 3 lines
This smells like it could be related to a problem that Ben Greear found recently (see "macvlan: Enable qdisc backoff logic"). When the hardware is busy, used to just drop the packet. With Ben's patch, we return
-EAGAIN
Quoted 8 lines
Quoted 4 lines
to qemu (or vhost-net) to trigger a resend.
I suppose what we really should do is feed that condition back to the guest network stack and implement the backoff in there.
Thanks for the pointer. I will take a look at this as I hadn't seen this patch earlier. Is there any way to figure out if this is the issue?
I think a good indication would be if this changes with/without the patch, and if you see -EAGAIN in qemu with the patch applied.
Arnd