Skip to contentSkip to conversation

Read I/O starvation with writeback RAID controller

DISCUSSION Martin Svec 4 participants 7 messages Feb 20, 2013 – Feb 22, 2013 SCSI

lore
Share & tools
lore
#0 lore raw

Hello,

I've noticed read I/O starvation problems of LIO iSCSI target when used on top of writeback-enabled HW RAID controller (PERC H700 with 1GB cache). For intensive mixed read-write workload in virtualized environments, writes are able to consume over 95% of the IOPS throughput and cause starvation of reads.

After a number of tests it seems to me it's a general issue of block layer I/O scheduling when running on top of a writeback device. If there is a write-intensive task, all writes go to the writeback cache with near-zero latency. This allows writer to quickly saturate the device with thousands of writes while using only a minimal fraction of queue depth. However, non-cached reads depend on spinning drive latencies which are orders of magnitude higher than writeback cache latencies, and so readers cannot submit so many requests per second as writers. Consequently, I guess the controller has totally wrong view of the incoming workload pattern, tries to satisfy the write flood first and the net result is inacceptable starvation of reads, with latencies up to hundreds of milliseconds.

A simple fio test with 1TiB block device where one thread does 4k random sync writes with iodepth=32 and one thread does 4k random reads with iodepth=32 shows that instead of the theoretical 50:50 IOPS ratio, the block device runs with 95:5 ratio in favor of writes. In fact, the imbalance is so high that even write iodepth=2 is enaugh to achieve the same numbers.

Real workloads that tend to exhibit this problem are: initial zeroing of a virtual machine disk, virtual machine migration, virtual machine cloning, intensive swapping of one virtual machine etc.

I tried to set WCE=1 on target iblock device, played with queue depths, tested all three I/O schedulers and their parameters, controller's parameters, but with no luck. To achieve reasonably good fairness, the only solution is to set nr_requests to 1 or disable controller's writeback cache at all -- at the expense of degraded overall performance :-(

Regarding nr_requests, there's obvious relation between iodepths and read starvation: if (nr_requests >= workload iodepth) then starvation surely occurs. Lowering nr_requests below this threshold slowly starts improving fairness and for every rd+wr iodepths pair, there exists sufficiently low nr_requests value at which IOPS ratio is finally balanced according to rd:wr iodepth ratio. Unfortunately it means there is no minimal nr_requests value suitable for all workloads. For iodepths around 2 to 8, only nr_requests=1 provides fair load balancing.

Is this a known problem? Does anybody find block layer parameters that elliminate this problem for iscsi-target storage in mixed random read-write environments like virtualization? Or should I start writing my own I/O scheduler? ;-)

Update: I've just found https://lkml.org/lkml/2012/12/10/550 (Read starvation by sync writes), where Jan Kara describes identical symptoms. But setting nr_requests=10000 doesn't help in my case. CC'ing LKML too (I'm not LKML subscriber).

Thanks,

Martin

#1 lore raw

Hi Martin,

CC'ing linux-scsi here, as aacraid doesn't have an official maintainer atm.

--nab

Quoted 66 lines · Martin Svec

Hello,

I've noticed read I/O starvation problems of LIO iSCSI target when used on top of writeback-enabled HW RAID controller (PERC H700 with 1GB cache). For intensive mixed read-write workload in virtualized environments, writes are able to consume over 95% of the IOPS throughput and cause starvation of reads.

After a number of tests it seems to me it's a general issue of block layer I/O scheduling when running on top of a writeback device. If there is a write-intensive task, all writes go to the writeback cache with near-zero latency. This allows writer to quickly saturate the device with thousands of writes while using only a minimal fraction of queue depth. However, non-cached reads depend on spinning drive latencies which are orders of magnitude higher than writeback cache latencies, and so readers cannot submit so many requests per second as writers. Consequently, I guess the controller has totally wrong view of the incoming workload pattern, tries to satisfy the write flood first and the net result is inacceptable starvation of reads, with latencies up to hundreds of milliseconds.

A simple fio test with 1TiB block device where one thread does 4k random sync writes with iodepth=32 and one thread does 4k random reads with iodepth=32 shows that instead of the theoretical 50:50 IOPS ratio, the block device runs with 95:5 ratio in favor of writes. In fact, the imbalance is so high that even write iodepth=2 is enaugh to achieve the same numbers.

Real workloads that tend to exhibit this problem are: initial zeroing of a virtual machine disk, virtual machine migration, virtual machine cloning, intensive swapping of one virtual machine etc.

I tried to set WCE=1 on target iblock device, played with queue depths, tested all three I/O schedulers and their parameters, controller's parameters, but with no luck. To achieve reasonably good fairness, the only solution is to set nr_requests to 1 or disable controller's writeback cache at all -- at the expense of degraded overall performance :-(

Regarding nr_requests, there's obvious relation between iodepths and read starvation: if (nr_requests >= workload iodepth) then starvation surely occurs. Lowering nr_requests below this threshold slowly starts improving fairness and for every rd+wr iodepths pair, there exists sufficiently low nr_requests value at which IOPS ratio is finally balanced according to rd:wr iodepth ratio. Unfortunately it means there is no minimal nr_requests value suitable for all workloads. For iodepths around 2 to 8, only nr_requests=1 provides fair load balancing.

Is this a known problem? Does anybody find block layer parameters that elliminate this problem for iscsi-target storage in mixed random read-write environments like virtualization? Or should I start writing my own I/O scheduler? ;-)

Update: I've just found https://lkml.org/lkml/2012/12/10/550 (Read starvation by sync writes), where Jan Kara describes identical symptoms. But setting nr_requests=10000 doesn't help in my case. CC'ing LKML too (I'm not LKML subscriber).

Thanks,

Martin

--
To unsubscribe from this list: send the line "unsubscribe target-devel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#2 lore raw

I'm sorry, I forgot to mention hardware details. It isn't aacraid, it is megaraid-based Dell PERC H700 w/ 1GB NVRAM and 12x 450GB 15k SAS drives in RAID-10. All in Dell R510 server.

Thanks,

Martin

Dne 20.2.2013 21:48, Nicholas A. Bellinger napsal(a):

Quoted 75 lines

Hi Martin,

CC'ing linux-scsi here, as aacraid doesn't have an official maintainer atm.

--nab

Quoted 66 lines · Martin Svec

Hello,

I've noticed read I/O starvation problems of LIO iSCSI target when used on top of writeback-enabled HW RAID controller (PERC H700 with 1GB cache). For intensive mixed read-write workload in virtualized environments, writes are able to consume over 95% of the IOPS throughput and cause starvation of reads.

After a number of tests it seems to me it's a general issue of block layer I/O scheduling when running on top of a writeback device. If there is a write-intensive task, all writes go to the writeback cache with near-zero latency. This allows writer to quickly saturate the device with thousands of writes while using only a minimal fraction of queue depth. However, non-cached reads depend on spinning drive latencies which are orders of magnitude higher than writeback cache latencies, and so readers cannot submit so many requests per second as writers. Consequently, I guess the controller has totally wrong view of the incoming workload pattern, tries to satisfy the write flood first and the net result is inacceptable starvation of reads, with latencies up to hundreds of milliseconds.

A simple fio test with 1TiB block device where one thread does 4k random sync writes with iodepth=32 and one thread does 4k random reads with iodepth=32 shows that instead of the theoretical 50:50 IOPS ratio, the block device runs with 95:5 ratio in favor of writes. In fact, the imbalance is so high that even write iodepth=2 is enaugh to achieve the same numbers.

Real workloads that tend to exhibit this problem are: initial zeroing of a virtual machine disk, virtual machine migration, virtual machine cloning, intensive swapping of one virtual machine etc.

I tried to set WCE=1 on target iblock device, played with queue depths, tested all three I/O schedulers and their parameters, controller's parameters, but with no luck. To achieve reasonably good fairness, the only solution is to set nr_requests to 1 or disable controller's writeback cache at all -- at the expense of degraded overall performance :-(

Regarding nr_requests, there's obvious relation between iodepths and read starvation: if (nr_requests >= workload iodepth) then starvation surely occurs. Lowering nr_requests below this threshold slowly starts improving fairness and for every rd+wr iodepths pair, there exists sufficiently low nr_requests value at which IOPS ratio is finally balanced according to rd:wr iodepth ratio. Unfortunately it means there is no minimal nr_requests value suitable for all workloads. For iodepths around 2 to 8, only nr_requests=1 provides fair load balancing.

Is this a known problem? Does anybody find block layer parameters that elliminate this problem for iscsi-target storage in mixed random read-write environments like virtualization? Or should I start writing my own I/O scheduler? ;-)

Update: I've just found https://lkml.org/lkml/2012/12/10/550 (Read starvation by sync writes), where Jan Kara describes identical symptoms. But setting nr_requests=10000 doesn't help in my case. CC'ing LKML too (I'm not LKML subscriber).

Thanks,

Martin

--
To unsubscribe from this list: send the line "unsubscribe target-devel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#3 lore raw

Hi Martin,

Quoted 4 lines · Martin Svec

I'm sorry, I forgot to mention hardware details. It isn't aacraid, it is megaraid-based Dell PERC H700 w/ 1GB NVRAM and 12x 450GB 15k SAS drives in RAID-10. All in Dell R510 server.

Jan Engelhardt (CC'ed) mentioned the currently out-of-tree ROW scheduler worked for him:

https://lkml.org/lkml/2012/12/11/534

Perhaps this would be worth a shot..?

--nab

Quoted 85 lines

Thanks,

Martin

Dne 20.2.2013 21:48, Nicholas A. Bellinger napsal(a):

Quoted 75 lines

Hi Martin,

CC'ing linux-scsi here, as aacraid doesn't have an official maintainer atm.

--nab

Quoted 66 lines · Martin Svec

Hello,

I've noticed read I/O starvation problems of LIO iSCSI target when used on top of writeback-enabled HW RAID controller (PERC H700 with 1GB cache). For intensive mixed read-write workload in virtualized environments, writes are able to consume over 95% of the IOPS throughput and cause starvation of reads.

After a number of tests it seems to me it's a general issue of block layer I/O scheduling when running on top of a writeback device. If there is a write-intensive task, all writes go to the writeback cache with near-zero latency. This allows writer to quickly saturate the device with thousands of writes while using only a minimal fraction of queue depth. However, non-cached reads depend on spinning drive latencies which are orders of magnitude higher than writeback cache latencies, and so readers cannot submit so many requests per second as writers. Consequently, I guess the controller has totally wrong view of the incoming workload pattern, tries to satisfy the write flood first and the net result is inacceptable starvation of reads, with latencies up to hundreds of milliseconds.

A simple fio test with 1TiB block device where one thread does 4k random sync writes with iodepth=32 and one thread does 4k random reads with iodepth=32 shows that instead of the theoretical 50:50 IOPS ratio, the block device runs with 95:5 ratio in favor of writes. In fact, the imbalance is so high that even write iodepth=2 is enaugh to achieve the same numbers.

Real workloads that tend to exhibit this problem are: initial zeroing of a virtual machine disk, virtual machine migration, virtual machine cloning, intensive swapping of one virtual machine etc.

I tried to set WCE=1 on target iblock device, played with queue depths, tested all three I/O schedulers and their parameters, controller's parameters, but with no luck. To achieve reasonably good fairness, the only solution is to set nr_requests to 1 or disable controller's writeback cache at all -- at the expense of degraded overall performance :-(

Regarding nr_requests, there's obvious relation between iodepths and read starvation: if (nr_requests >= workload iodepth) then starvation surely occurs. Lowering nr_requests below this threshold slowly starts improving fairness and for every rd+wr iodepths pair, there exists sufficiently low nr_requests value at which IOPS ratio is finally balanced according to rd:wr iodepth ratio. Unfortunately it means there is no minimal nr_requests value suitable for all workloads. For iodepths around 2 to 8, only nr_requests=1 provides fair load balancing.

Is this a known problem? Does anybody find block layer parameters that elliminate this problem for iscsi-target storage in mixed random read-write environments like virtualization? Or should I start writing my own I/O scheduler? ;-)

Update: I've just found https://lkml.org/lkml/2012/12/10/550 (Read starvation by sync writes), where Jan Kara describes identical symptoms. But setting nr_requests=10000 doesn't help in my case. CC'ing LKML too (I'm not LKML subscriber).

Thanks,

Martin

--
To unsubscribe from this list: send the line "unsubscribe target-devel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe target-devel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#4 lore raw

Hi,

Dne 21.2.2013 23:01, Nicholas A. Bellinger napsal(a):

Quoted 8 lines

Hi Martin,

Quoted 4 lines · Martin Svec

I'm sorry, I forgot to mention hardware details. It isn't aacraid, it is megaraid-based Dell PERC H700 w/ 1GB NVRAM and 12x 450GB 15k SAS drives in RAID-10. All in Dell R510 server.

Quoted 6 lines

Jan Engelhardt (CC'ed) mentioned the currently out-of-tree ROW scheduler worked for him:

https://lkml.org/lkml/2012/12/11/534

Perhaps this would be worth a shot..?

Yes, I've already tried the ROW scheduler. It helped for some low iodepths depending on quantum settings but generally didn't solve the problem. I think the key issue is that none of the schedulers can throttle I/O according to e.g. average request roundtrip time. Shaohua Li is right here: https://lkml.org/lkml/2012/12/11/598 -- as long as there's free room in device's queue they blindly dispatch requests to it.

Which is exactly what I see in deadline scheduler fifo queues: There're no read requests to be scheduled between writes because all readers are starving. So the scheduler keeps dispatching writes using all the remaining capacity of device queue. Which in turn worses the read starvation. Bigger queue depth and bigger writeback cache means higher chance for read starvation even from a single writer.

Thanks,
Martin

Quoted 89 lines

--nab

Quoted 85 lines

Thanks,

Martin

Dne 20.2.2013 21:48, Nicholas A. Bellinger napsal(a):

Quoted 75 lines

Hi Martin,

CC'ing linux-scsi here, as aacraid doesn't have an official maintainer atm.

--nab

Quoted 66 lines · Martin Svec

Hello,

I've noticed read I/O starvation problems of LIO iSCSI target when used on top of writeback-enabled HW RAID controller (PERC H700 with 1GB cache). For intensive mixed read-write workload in virtualized environments, writes are able to consume over 95% of the IOPS throughput and cause starvation of reads.

After a number of tests it seems to me it's a general issue of block layer I/O scheduling when running on top of a writeback device. If there is a write-intensive task, all writes go to the writeback cache with near-zero latency. This allows writer to quickly saturate the device with thousands of writes while using only a minimal fraction of queue depth. However, non-cached reads depend on spinning drive latencies which are orders of magnitude higher than writeback cache latencies, and so readers cannot submit so many requests per second as writers. Consequently, I guess the controller has totally wrong view of the incoming workload pattern, tries to satisfy the write flood first and the net result is inacceptable starvation of reads, with latencies up to hundreds of milliseconds.

A simple fio test with 1TiB block device where one thread does 4k random sync writes with iodepth=32 and one thread does 4k random reads with iodepth=32 shows that instead of the theoretical 50:50 IOPS ratio, the block device runs with 95:5 ratio in favor of writes. In fact, the imbalance is so high that even write iodepth=2 is enaugh to achieve the same numbers.

Real workloads that tend to exhibit this problem are: initial zeroing of a virtual machine disk, virtual machine migration, virtual machine cloning, intensive swapping of one virtual machine etc.

I tried to set WCE=1 on target iblock device, played with queue depths, tested all three I/O schedulers and their parameters, controller's parameters, but with no luck. To achieve reasonably good fairness, the only solution is to set nr_requests to 1 or disable controller's writeback cache at all -- at the expense of degraded overall performance :-(

Regarding nr_requests, there's obvious relation between iodepths and read starvation: if (nr_requests>= workload iodepth) then starvation surely occurs. Lowering nr_requests below this threshold slowly starts improving fairness and for every rd+wr iodepths pair, there exists sufficiently low nr_requests value at which IOPS ratio is finally balanced according to rd:wr iodepth ratio. Unfortunately it means there is no minimal nr_requests value suitable for all workloads. For iodepths around 2 to 8, only nr_requests=1 provides fair load balancing.

Is this a known problem? Does anybody find block layer parameters that elliminate this problem for iscsi-target storage in mixed random read-write environments like virtualization? Or should I start writing my own I/O scheduler? ;-)

Update: I've just found https://lkml.org/lkml/2012/12/10/550 (Read starvation by sync writes), where Jan Kara describes identical symptoms. But setting nr_requests=10000 doesn't help in my case. CC'ing LKML too (I'm not LKML subscriber).

Thanks,

Martin

--
To unsubscribe from this list: send the line "unsubscribe target-devel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe target-devel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#5 lore raw
Quoted 8 lines · Martin Svec

Yes, I've already tried the ROW scheduler. It helped for some low iodepths depending on quantum settings but generally didn't solve the problem. I think the key issue is that none of the schedulers can throttle I/O according to e.g. average request roundtrip time. Shaohua Li is right here:
https://lkml.org/lkml/2012/12/11/598 -- as long as there's free room in device's queue they blindly dispatch requests to it.

Quoted 6 lines

Which is exactly what I see in deadline scheduler fifo queues: There're no read requests to be scheduled between writes because all readers are starving. So the scheduler keeps dispatching writes using all the remaining capacity of device queue. Which in turn worses the read starvation. Bigger queue depth and bigger writeback cache means higher chance for read starvation even from a single writer.

Sounds just like the bufferbloat problem in networking. Waiting for codel for the block layer :)

#6 lore raw
Quoted 17 lines · Jan Engelhardt
Quoted 8 lines · Martin Svec

Yes, I've already tried the ROW scheduler. It helped for some low iodepths depending on quantum settings but generally didn't solve the problem. I think the key issue is that none of the schedulers can throttle I/O according to e.g. average request roundtrip time. Shaohua Li is right here:
https://lkml.org/lkml/2012/12/11/598 -- as long as there's free room in device's queue they blindly dispatch requests to it.

Quoted 6 lines

Which is exactly what I see in deadline scheduler fifo queues: There're no read requests to be scheduled between writes because all readers are starving. So the scheduler keeps dispatching writes using all the remaining capacity of device queue. Which in turn worses the read starvation. Bigger queue depth and bigger writeback cache means higher chance for read starvation even from a single writer.

Quoted 2 lines

Sounds just like the bufferbloat problem in networking. Waiting for codel for the block layer :)

Is there any way to somehow have the reads jump to the head of the queue in the disk controller?

Otherwise it seems like we might need to minimize the disk cache usage and do the scheduling in software.

This effectively mirrors what the codel people are doing with using tiny tx ring buffers to fight bufferbloat. The difference is that with a NIC all you have to do is make sure the buffer doesn't empty and you get full speed whereas with a disk the more you stuff in the cache the better it can schedule things.

Chris