ARP table question

I have 500 mac-vlans on a system talking to 500 other mac-vlans. My problem is that the arp-table gets extremely huge because every time an arp-request comes in on all mac-vlans, a stale arp entry is added for each mac-vlan. I have filtering turned on, but that doesn't help because the neigh_event_ns call below will cause a stale neighbor entry to be created regardless of whether a replay will be sent or not.

Maybe the neigh_event code should be below the checks for dont_send, and only create check neigh_event_ns if we are !dont_send?

Quoted 9 lines · Ben Greear

I have 500 mac-vlans on a system talking to 500 other mac-vlans. My problem is that the arp-table gets extremely huge because every time an arp-request comes in on all mac-vlans, a stale arp entry is added for each mac-vlan. I have filtering turned on, but that doesn't help because the neigh_event_ns call below will cause a stale neighbor entry to be created regardless of whether a replay will be sent or not.
Maybe the neigh_event code should be below the checks for dont_send, and only create check neigh_event_ns if we are !dont_send?

The attached patch makes it work much better for me. The patch will cause the code to NOT create a stale neighbor entry if we are not going to respond to the ARP request. The old code *would* create a stale entry even if we are not going to respond.

This is against 2.6.25.15.

Signed-off-by Ben Greear<[email protected]>

This change makes a lot of sense to me, I'll add it to net-next-2.6 so it can cook in there for a while just in case there are some unwanted side-effects.

Quoted 3 lines · David Miller

This change makes a lot of sense to me, I'll add it to net-next-2.6 so it can cook in there for a while just in case there are some unwanted side-effects.

Thanks Dave.

I think I found another problem as well: If I start 1 TCP and 1 UDP connection between each of the 500 interfaces on mac-vlans, the ARP tables will not converge.

It seems to be because mac-vlan has to copy broadcast packets to every mac-vlan on a physical device, there are just too many packets:

500 vlans arping once per second means 500 pkts per second on the other NIC.
Other NIC must copy these 500 times,
so, 250000 packets per second in each direction are processed by the stack (they are not all on the wire, at least).

A few get through and those UDP/TCP connections start consuming bandwidth, which clogs up the 1G link enough that other responses are lost most of the time.

I'm going to try to work on some sort of random backoff for ARP that can be enabled in this situation next.

Ok, here is the patch that implements this. The idea is to spread out arp requests when you do something like start 500 TCP connections on 500 MAC-VLANs talking to 500 other MAC-VLANs.

With a retrans timer of 1 sec, and a high volume of traffic, and a semi flaky network in between, my system will not resolve the ARPs and the retransmits overload my processors.

Setting the retrans timer to 5 secs on my system also works, so I'm not sure if this patch is really required, but it might help keep arp requests somewhat random in cases where arp timers would otherwise try to all fire at the same time.

This is against 2.6.25.20 plus my patches, but I believe it should apply to a clean 2.6.25.20 as well.

Comments are welcome.

Signed-Off-By Ben Greear<[email protected]>

Thanks,
Ben

------------------------------------------------------------------------

Documentation/filesystems/proc.txt +10 -0

index 518ebe6..4c805b3 100644
@@ -2028,6 +2028,16 @@ Expression of retrans_time, which is deprecated, is in 1/100 seconds (for
 IPv4) or in jiffies (for IPv6).
 Expression of retrans_time_ms is in milliseconds.
 
+
+retrans_rand_backof_ms
+----------------------
+
+This is an extra delay (ms) for the retransmit timer.  A random value between
+0 and retrans_rand_backof_ms will be added to the retrans_timer.  Default
+is zero.  Setting this to a larger value will help large broadcast domains
+resolve ARP (for instance, 500 mac-vlans talking to 500 other mac-vlans).
+
+
 unres_qlen
 ----------
...

net/core/neighbour.c +0 -0

index 19b8e00..ec1f048 100644
@@ -765,6 +765,13 @@ static __inline__ int neigh_max_probes(struct neighbour *n)
 		p->ucast_probes + p->app_probes + p->mcast_probes);

net/core/neighbour.c static __inline__ int neigh_max_probes(struct neighbour *n) · line 775

Quoted 8 lines

net/core/neighbour.c +7 -0

+static unsigned long neigh_rand_retry(struct neighbour* neigh) {
+    if (neigh->parms->retrans_rand_backoff) {
+        return net_random() % neigh->parms->retrans_rand_backoff;
+    }
+    return 0;
+}
+
 /* Called when a timer expires for a neighbour entry. */

I thought that mod was something we tried to avoid? Could you instead use something that isn't random but perhaps varies among all the requests? Say some of the low-order bits of the IP being resolved?

Quoted 11 lines · Rick Jones

net/core/neighbour.c static __inline__ int neigh_max_probes(struct neighbour *n) · line 775

Quoted 8 lines

net/core/neighbour.c +7 -0

+static unsigned long neigh_rand_retry(struct neighbour* neigh) {
+    if (neigh->parms->retrans_rand_backoff) {
+        return net_random() % neigh->parms->retrans_rand_backoff;
+    }
+    return 0;
+}
+
 /* Called when a timer expires for a neighbour entry. */

I thought that mod was something we tried to avoid? Could you instead use something that isn't random but perhaps varies among all the requests? Say some of the low-order bits of the IP being resolved?

This is only called when we are going to retransmit an ARP, which shouldn't be in any sort of hot path, so I figured MOD was fine.

The net_random is a very cheap method (last I checked), as well.

So, I think that part is OK as it is, but I'm open to persuasion :)

Quoted 18 lines · Ben Greear

Quoted 11 lines · Rick Jones

net/core/neighbour.c static __inline__ int neigh_max_probes(struct neighbour *n) · line 775

Quoted 8 lines

net/core/neighbour.c +7 -0

+static unsigned long neigh_rand_retry(struct neighbour* neigh) {
+    if (neigh->parms->retrans_rand_backoff) {
+        return net_random() % neigh->parms->retrans_rand_backoff;
+    }
+    return 0;
+}
+
 /* Called when a timer expires for a neighbour entry. */

I thought that mod was something we tried to avoid? Could you instead use something that isn't random but perhaps varies among all the requests? Say some of the low-order bits of the IP being resolved?

This is only called when we are going to retransmit an ARP, which shouldn't be in any sort of hot path, so I figured MOD was fine.

Quoted 4 lines

The net_random is a very cheap method (last I checked), as well.

So, I think that part is OK as it is, but I'm open to persuasion :)

Perhaps I'm confused, or simply channeling Emily Litella again, but if you only do this on the 1st through Nth retransmissions (ie after the first retransmission timer has popped) don't you still have a thundering herd problem on the first transmission and the first retransmission of ARP requests?

Quoted 16 lines · Rick Jones

Quoted 15 lines · Ben Greear

Quoted 11 lines · Rick Jones

net/core/neighbour.c static __inline__ int neigh_max_probes(struct neighbour *n) · line 775

Quoted 8 lines

net/core/neighbour.c +7 -0

+static unsigned long neigh_rand_retry(struct neighbour* neigh) {
+    if (neigh->parms->retrans_rand_backoff) {
+        return net_random() % neigh->parms->retrans_rand_backoff;
+    }
+    return 0;
+}
+
 /* Called when a timer expires for a neighbour entry. */

I thought that mod was something we tried to avoid? Could you instead use something that isn't random but perhaps varies among all the requests? Say some of the low-order bits of the IP being resolved?

Quoted 8 lines

Quoted 7 lines

This is only called when we are going to retransmit an ARP, which shouldn't be in any sort of hot path, so I figured MOD was fine.

The net_random is a very cheap method (last I checked), as well.

So, I think that part is OK as it is, but I'm open to persuasion :)

Perhaps I'm confused, or simply channeling Emily Litella again, but if you only do this on the 1st through Nth retransmissions (ie after the first retransmission timer has popped) don't you still have a thundering herd problem on the first transmission and the first retransmission of ARP requests?

You'd certainly have it on the first transmission, but I think from there on the randomness should kick in. This is a pretty rare case, and I'd rather not slow down the initial ARP. If we *are* in the overload situation, then the network can just purge/drop/whatever the initial flood and then the retransmits should start doing their random thing. On my system, it still takes maybe 30 seconds for all the ARPs to resolve since a good deal of the requests and/or responses are being lost.

After some more testing, I can still get it into a bad state if I have a retrans timer of 1 sec and a randomness of 5 secs and manage to cause all 1000 arp entries to go stale at once (by yanking a cable, for instance).

It seems I have to bump up the base timer to 3-5 seconds (I'm leaving the random backoff at 5 secs as well).

Quoted 7 lines

After some more testing, I can still get it into a bad state if I have a retrans timer of 1 sec and a randomness of 5 secs and manage to cause all 1000 arp entries to go stale at once (by yanking a cable, for instance).

It seems I have to bump up the base timer to 3-5 seconds (I'm leaving the random backoff at 5 secs as well).

This scheme still seems hackish to me, so I'm going to defer on this for now.

You think something like an exponential backoff capped at some user-configurable
max-value would be better?