+++ This bug was initially created as a clone of Bug #1665029 +++ Description of problem: A large-file sequential read test reports better throughput with read-ahead and io-cache xlators turned off. One test is obviously not enough to decide the fate of these xlators. But through this bz we can answer some of the relevant questions. Version-Release number of selected component (if applicable): glusterfs-*5.2-1.el7.x86_64 kernel-3.10.0-957.el7.x86_64 (RHEL 7.6) How reproducible: Consistently Steps to Reproduce: 1. fio write test to generate data set: fio --name=initialwrite --ioengine=sync --rw=write \ --direct=0 --create_on_open=1 --end_fsync=1 --bs=128k \ --directory=/mnt/glustervol/ --filename_format=f.\$jobnum.\$filenum \ --filesize=16g --size=16g --numjobs=4 2. unmount and re-mount volume on client 3. fio sequential read test to read back the data written in step 1: fio --name=readtest --ioengine=sync --rw=read \ --direct=0 --bs=128k --directory=/mnt/glustervol/ \ --filename_format=f.\$jobnum.\$filenum --filesize=16g \ --size=16g --numjobs=4 Actual results: With default volume settings: READ: bw=581MiB/s (610MB/s), 145MiB/s-146MiB/s (152MB/s-153MB/s), io=64.0GiB (68.7GB), run=112401-112717msec With read-ahead and io-cache turned off: READ: bw=1083MiB/s (1136MB/s), 271MiB/s-271MiB/s (284MB/s-284MB/s), io=64.0GiB (68.7GB), run=60487-60491msec So in this test, there is a significant performance gain with these xlators turned off. The low performance with default volume settings was first seen in another bz:https://bugzilla.redhat.com/show_bug.cgi?id=1664934#c0. Expected results: performance xlators should not degrade performance Additional info: --- Additional comment from Manoj Pillai on 2019-01-10 13:18:03 UTC --- Data showing that both read-ahead and io-cache cause performance degradation. I'm modifying the test in comment #0 here, substituting numjobs=2, instead of 4. Test sequence: fio --name=initialwrite --ioengine=sync --rw=write --direct=0 --create_on_open=1 --end_fsync=1 --bs=128k --directory=/mnt/glustervol/ --filename_format=f.\$jobnum.\$filenum --filesize=16g --size=16g --numjobs=2 [unmount and mount volume] fio --name=readtest --ioengine=sync --rw=read --direct=0 --bs=128k --directory=/mnt/glustervol/ --filename_format=f.\$jobnum.\$filenum --filesize=16g --size=16g --numjobs=2 Result with default settings: READ: bw=485MiB/s (509MB/s), 243MiB/s-243MiB/s (254MB/s-255MB/s), io=32.0GiB (34.4GB), run=67504-67522msec Result with read-ahead turned off: READ: bw=776MiB/s (813MB/s), 388MiB/s-388MiB/s (407MB/s-407MB/s), io=32.0GiB (34.4GB), run=42220-42237msec Result with read-ahead and io-cache turned off: READ: bw=1108MiB/s (1162MB/s), 554MiB/s-554MiB/s (581MB/s-581MB/s), io=32.0GiB (34.4GB), run=29565-29573msec --- Additional comment from Raghavendra G on 2019-01-11 06:40:57 UTC --- Some observations while debugging the performance degradation with gluster read-ahead: * Kernel too does read-ahead and it sends parallel read-requests as part of this. * client-io-threads is on in this configuration. The above two points mean parallel requests sent by kernel can reach read-ahead out of order. This means read-ahead no longer sees read requests at sequential contiguous offsets and hence it things reads are random. For random reads, it resets the read sequence. But when requests reach read-ahead in order, read-ahead is turned on again. Due to this intermittent toggling, much of read-ahead data is wasted regressing the performance. With client-io-threads off and I can no longer see the regression for the test case given. If I run the test with single fio job (--numjobs=1), gluster read-ahead on outperforms gluster read-ahead off on my setup. [1] https://review.gluster.org/#/c/glusterfs/+/20981/ --- Additional comment from Raghavendra G on 2019-01-11 06:42:55 UTC --- (In reply to Raghavendra G from comment #2) > Some observations while debugging the performance degradation with gluster > read-ahead: > > * Kernel too does read-ahead and it sends parallel read-requests as part of > this. > * client-io-threads is on in this configuration. > > The above two points mean parallel requests sent by kernel can reach > read-ahead out of order. This means read-ahead no longer sees read requests > at sequential contiguous offsets and hence it things reads are random. For > random reads, it resets the read sequence. But when requests reach > read-ahead in order, read-ahead is turned on again. Due to this intermittent > toggling, much of read-ahead data is wasted regressing the performance. With > client-io-threads off and I can no longer see the regression for the test > case given. If I run the test with single fio job (--numjobs=1), gluster > read-ahead on outperforms gluster read-ahead off on my setup. ... single fio job (--numjobs=1), gluster read-ahead on with client-io-threads off outperforms gluster read-ahead off with client-io-threads off. > > [1] https://review.gluster.org/#/c/glusterfs/+/20981/ --- Additional comment from Manoj Pillai on 2019-01-11 12:40:52 UTC --- (In reply to Manoj Pillai from comment #1) > Data showing that both read-ahead and io-cache cause performance > degradation. I'm modifying the test in comment #0 here, substituting > numjobs=2, instead of 4. > > Test sequence: > fio --name=initialwrite --ioengine=sync --rw=write --direct=0 > --create_on_open=1 --end_fsync=1 --bs=128k --directory=/mnt/glustervol/ > --filename_format=f.\$jobnum.\$filenum --filesize=16g --size=16g --numjobs=2 > [unmount and mount volume] > fio --name=readtest --ioengine=sync --rw=read --direct=0 --bs=128k > --directory=/mnt/glustervol/ --filename_format=f.\$jobnum.\$filenum > --filesize=16g --size=16g --numjobs=2 > > Result with default settings: > READ: bw=485MiB/s (509MB/s), 243MiB/s-243MiB/s (254MB/s-255MB/s), io=32.0GiB > (34.4GB), run=67504-67522msec > > Result with read-ahead turned off: > READ: bw=776MiB/s (813MB/s), 388MiB/s-388MiB/s (407MB/s-407MB/s), io=32.0GiB > (34.4GB), run=42220-42237msec > > Result with read-ahead and io-cache turned off: > READ: bw=1108MiB/s (1162MB/s), 554MiB/s-554MiB/s (581MB/s-581MB/s), > io=32.0GiB (34.4GB), run=29565-29573msec Result with ciot=off, io-cache=off, gluster ra=on, read-ahead-page-count=10: [these settings are based on comment #2 and comment #3] READ: bw=975MiB/s (1023MB/s), 488MiB/s-488MiB/s (511MB/s-512MB/s), io=32.0GiB (34.4GB) Comparing the best results seen (1108 vs 975), the gluster ra=off case is still a little bit better. Result with ciot=off, io-cache=on, gluster ra=on, read-ahead-page-count=10: READ: bw=674MiB/s (706MB/s), 337MiB/s-339MiB/s (353MB/s-355MB/s), io=32.0GiB (34.4GB) --- Additional comment from Amar Tumballi on 2019-02-12 12:11:51 UTC --- Please go ahead and disable it by default in upstream master. Lets get a run done with these values, and if the performance is good without these 2 translators, then we can backport the patch to glusterfs-6 branch. Otherwise, it would allow us another 2 months to validate it in upstream master before glusterfs-7.
REVIEW: https://review.gluster.org/22203 (read-ahead/io-cache: turn off by default) posted (#1) for review on master by Raghavendra G
*** Bug 1094328 has been marked as a duplicate of this bug. ***
*** Bug 1665029 has been marked as a duplicate of this bug. ***
REVIEW: https://review.gluster.org/22203 (read-ahead/io-cache: turn off by default) merged (#4) on master by Raghavendra G