Apologies. Pressed 'send' even before I was done.
On Tue, Jun 20, 2017 at 11:39 AM, Krutika Dhananjay <kdhananj@xxxxxxxxxx> wrote:
The disk utilization didn't improve.∆ (2,3) is down to 804us.Avg READ latency as seen by the io-stats translator that sits above client-io-threads came down to 963us (from 1666us).The IOPs increased to ~50K (from 38K).Some update on this topic:I ran fio again, this time with Raghavendra's epoll-rearm patch @ https://review.gluster.org/17391
From code reading, it appears there is some serialization between POLLIN, POLLOUT and POLLERR events for a given socket because of socket_private->lock which they all contend for.
Discussed the same with Raghavendra G.
(I think he already alluded to the same point in this thread earlier.)
Let me make some quick dirty changes to see if fixing this serialization improves performance further and I'll update the thread accordingly.
(I think he already alluded to the same point in this thread earlier.)
Let me make some quick dirty changes to see if fixing this serialization improves performance further and I'll update the thread accordingly.
-Krutika
On Sat, Jun 10, 2017 at 12:47 AM, Manoj Pillai <mpillai@xxxxxxxxxx> wrote:So comparing the key latency, ∆ (2,3), in the two cases:iodepth=1: 171 usiodepth=8: 1453 us (in the ballpark of 171*8=1368). That's not good! (I wonder if that relation roughly holds up for other values of iodepth).This data doesn't conclusively establish that the problem is in gluster. You'd see similar results if the network were saturated, like Vijay suggested. But from what I remember of this test, the throughput here is far too low for that to be the case.-- ManojOn Thu, Jun 8, 2017 at 6:37 PM, Krutika Dhananjay <kdhananj@xxxxxxxxxx> wrote:Indeed the latency on the client side dropped with iodepth=1. :)I ran the test twice and the results were consistent.Here are the exact numbers:
Translator Position Avg Latency of READ fop as seen by this translator1. parent of client-io-threads437us ∆ (1,2) = 69us2. parent of protocol/client-0 368us∆ (2,3) = 171us----------------- end of client stack -------------------------------------- beginning of brick stack --------------3. child of protocol/server197us ∆ (3,4) = 4us4. parent of io-threads193us ∆ (4,5) = 32us5. child-of-io-threads161us ∆ (5,6) = 11us6. parent of storage/posix150us
...---------------- end of brick stack ------------------------Will continue reading code and get back when I find sth concrete.-KrutikaOn Thu, Jun 8, 2017 at 12:22 PM, Manoj Pillai <mpillai@xxxxxxxxxx> wrote:Thanks. So I was suggesting a repeat of the test but this time with iodepth=1 in the fio job. If reducing the no. of concurrent requests reduces drastically the high latency you're seeing from the client-side, that would strengthen the hypothesis than serialization/contention among concurrent requests at the n/w layers is the root cause here.-- ManojOn Thu, Jun 8, 2017 at 11:46 AM, Krutika Dhananjay <kdhananj@xxxxxxxxxx> wrote:-KrutikaI have 3 vms reading from one mount, and each of these vms is running the above job in parallel.Hi,This is what my job file contains:
[global]
ioengine=libaio
#unified_rw_reporting=1
randrepeat=1
norandommap=1
group_reporting
direct=1
runtime=60
thread
size=16g
[workload]
bs=4k
rw=randread
iodepth=8
numjobs=1
file_service_type=random
filename=/perf5/iotest/fio_5
filename=/perf6/iotest/fio_6
filename=/perf7/iotest/fio_7
filename=/perf8/iotest/fio_8On Tue, Jun 6, 2017 at 9:14 PM, Manoj Pillai <mpillai@xxxxxxxxxx> wrote:On Tue, Jun 6, 2017 at 5:05 PM, Krutika Dhananjay <kdhananj@xxxxxxxxxx> wrote:Based on a 60-second run of randrd test and subsequent analysis of the stats dumped by the individual io-stats instances, the following is what I found:On the brick stack: Below protocol/server, above and below io-threads and just above storage/posix.On the client stack: Above client-io-threads and above protocol/client-0 (the first child of AFR).3 node cluster; 1x3 plain replicate volume with group virt settings, direct-io.Before I get to the results, a little bit about the configuration ...Hi,As part of identifying performance bottlenecks within gluster stack for VM image store use-case, I loaded io-stats at multiple points on the client and brick stack and ran randrd test using fio from within the hosted vms in parallel.3 FUSE clients, one per node in the cluster (which implies reads are served from the replica that is local to the client).io-stats was loaded at the following places:Translator Position Avg Latency of READ fop as seen by this translator1. parent of client-io-threads1666us ∆ (1,2) = 50us2. parent of protocol/client-0 1616us∆ (2,3) = 1453us----------------- end of client stack -------------------------------------- beginning of brick stack -----------3. child of protocol/server163us ∆ (3,4) = 7us4. parent of io-threads156us ∆ (4,5) = 20us5. child-of-io-threads136us ∆ (5,6) = 11us6. parent of storage/posix125us
...---------------- end of brick stack ------------------------So it seems like the biggest bottleneck here is a combination of the network + epoll, rpc layer?I must admit I am no expert with networks, but I'm assuming if the client is reading from the local brick, theneven latency contribution from the actual network won't be much, in which case bulk of the latency is coming from epoll, rpc layer, etc at both client and brick end? Please correct me if I'm wrong.I will, of course, do some more runs and confirm if the pattern is consistent.-KrutikaReally interesting numbers! How many concurrent requests are in flight in this test? Could you post the fio job? I'm wondering if/how these latency numbers change if you reduce the number of concurrent requests.-- Manoj
_______________________________________________ Gluster-devel mailing list Gluster-devel@xxxxxxxxxxx http://lists.gluster.org/mailman/listinfo/gluster-devel
