Hi, Kumaran,that would explain the smaller IOs before the reboot, but not the larger-than-4MiB IOs afterwards on that machine.
Then, I already saw that the numaMemoryInterleave setting seems to have no effect (on that very installation), I just have not yet requested a PMR for it. I'd checked memory usage of course and saw that regardless of this setting always one socket's memory is almost completely consumed while the other one's is rather empty - looks like a bug to me, but that needs further investigation.
Uwe On 24.02.22 15:32, Kumaran Rajaram wrote:
Hi Uwe,>> But what puzzles me even more: one of the server compiles IOs even smaller, varying between 3.2MiB and 3.6MiB mostly - both for reads and writes ... I just cannot see why.IMHO, If GPFS on this particular NSD server was restarted often during the setup, then it is possible that the GPFS pagepool may not be contiguous. As a result, GPFS 8MiB buffer in the pagepool might be a scatter-gather (SG) list with many small entries (in the memory) resulting in smaller I/O when these buffers are issued to the disks. The fix would be to reboot the server and start GPFS so that pagepool is contiguous resulting in 8MiB buffer to be comprised of 1 (or fewer) SG entries.>>In the current situation (i.e. with IOs bit larger than 4MiB) setting max_sectors_kB to 4096 might do the trick, but as I do not know the cause for that behaviour it might well start to issue IOs >>smaller than 4MiB again at some point, so that is not a nice solution.It will be advised not to restart GPFS often in the NSD servers (in production) to keep the pagepool contiguous. Ensure that there is enough free memory in NSD server and not run any memory intensive jobs so that pagepool is not impacted (e.g. swapped out).Also, enable GPFS numaMemoryInterleave=yes and verify that pagepool is equally distributed across the NUMA domains for good performance. GPFS numaMemoryInterleave=yes requires that numactl packages are installed and then GPFS restarted.# mmfsadm dump config | egrep "numaMemory|pagepool " ! numaMemoryInterleave yes ! pagepool 282394099712 # pgrep mmfsd | xargs numastat -p Per-node process memory usage (in MBs) for PID 2120821 (mmfsd) Node 0 Node 1 Total --------------- --------------- --------------- Huge 0.00 0.00 0.00 Heap 1.26 3.26 4.52 Stack 0.01 0.01 0.02 Private 137710.43 137709.96 275420.39 ---------------- --------------- --------------- --------------- Total 137711.70 137713.23 275424.92 My two cents, -Kums Kumaran Rajaram*From:* gpfsug-discuss-boun...@spectrumscale.org <gpfsug-discuss-boun...@spectrumscale.org> *On Behalf Of *Uwe Falke*Sent:* Wednesday, February 23, 2022 8:04 PM *To:* gpfsug-discuss@spectrumscale.org *Subject:* Re: [gpfsug-discuss] IO sizes Hi,the test bench is gpfsperf running on up to 12 clients with 1...64 threads doing sequential reads and writes , file size per gpfsperf process is 12TB (with 6TB I saw caching effects in particular for large thread numbers ...)As I wrote initially: GPFS is issuing nothing but 8MiB IOs to the data disks, as expected in that case.Interesting thing though:I have rebooted the suspicious node. Now, it does not issue smaller IOs than the others, but -- unbelievable -- larger ones (up to about 4.7MiB). This is still harmful as also that size is incompatible with full stripe writes on the storage ( 8+2 disk groups, i.e. logically RAID6)Currently, I draw this information from the storage boxes; I have not yet checked iostat data for that benchmark test after the reboot (before, when IO sizes were smaller, we saw that both in iostat and in the perf data retrieved from the storage controllers).And: we have a separate data pool , hence dataOnly NSDs, I am just talking about these ...As for "Are you sure that Linux OS is configured the same on all 4 NSD servers?." - of course there are not two boxes identical in the world. I have actually not installed those machines, and, yes, i also considered reinstalling them (or at least the disturbing one).However, I do not have reason to assume or expect a difference, the supplier has just implemented these systems recently from scratch.In the current situation (i.e. with IOs bit larger than 4MiB) setting max_sectors_kB to 4096 might do the trick, but as I do not know the cause for that behaviour it might well start to issue IOs smaller than 4MiB again at some point, so that is not a nice solution.Thanks Uwe On 23.02.22 22:20, Andrew Beattie wrote: Alex, Metadata will be 4Kib Depending on the filesystem version you will also have subblocks to consider V4 filesystems have 1/32 subblocks, V5 filesystems have 1/1024 subblocks (assuming metadata and data block size is the same) My first question would be is “ Are you sure that Linux OS is configured the same on all 4 NSD servers?. My second question would be do you know what your average file size is if most of your files are smaller than your filesystem block size, then you are always going to be performing writes using groups of subblocks rather than a full block writes. Regards, Andrew On 24 Feb 2022, at 04:39, Alex Chekholko <a...@calicolabs.com> <mailto:a...@calicolabs.com> wrote: Hi, Metadata I/Os will always be smaller than the usual data block size, right? Which version of GPFS? Regards, Alex On Wed, Feb 23, 2022 at 10:26 AM Uwe Falke <uwe.fa...@kit.edu> <mailto:uwe.fa...@kit.edu> wrote: Dear all, sorry for asking a question which seems ZjQcmQRYFpfptBannerStart This Message Is From an External Sender This message came from outside your organization. ZjQcmQRYFpfptBannerEnd Hi, Metadata I/Os will always be smaller than the usual data block size, right? Which version of GPFS? Regards, Alex On Wed, Feb 23, 2022 at 10:26 AM Uwe Falke <uwe.fa...@kit.edu> wrote: Dear all, sorry for asking a question which seems not directly GPFS related: In a setup with 4 NSD servers (old-style, with storage controllers in the back end), 12 clients and 10 Seagate storage systems, I do see in benchmark tests that just one of the NSD servers does send smaller IO requests to the storage than the other 3 (that is, both reads and writes are smaller). The NSD servers form 2 pairs, each pair is connected to 5 seagate boxes ( one server to the controllers A, the other one to controllers B of the Seagates, resp.). All 4 NSD servers are set up similarly: kernel: 3.10.0-1160.el7.x86_64 #1 SMP HBA: Broadcom / LSI Fusion-MPT 12GSAS/PCIe Secure SAS38xx driver : mpt3sas 31.100.01.00 max_sectors_kb=8192 (max_hw_sectors_kb=16383 , not 16384, as limited by mpt3sas) for all sd devices and all multipath (dm) devices built on top. scheduler: deadline multipath (actually we do have 3 paths to each volume, so there is some asymmetry, but that should not affect the IOs, shouldn't it?, and if it did we would see the same effect in both pairs of NSD servers, but we do not). All 4 storage systems are also configured the same way (2 disk groups / pools / declustered arrays, one managed by ctrl A, one by ctrl B, and 8 volumes out of each; makes altogether 2 x 8 x 10 = 160 NSDs). GPFS BS is 8MiB , according to iohistory (mmdiag) we do see clean IO requests of 16384 disk blocks (i.e. 8192kiB) from GPFS. The first question I have - but that is not my main one: I do see, both in iostat and on the storage systems, that the default IO requests are about 4MiB, not 8MiB as I'd expect from above settings (max_sectors_kb is really in terms of kiB, not sectors, cf. https://www.kernel.org/doc/Documentation/block/queue-sysfs.txt <https://nam02.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.kernel.org%2Fdoc%2FDocumentation%2Fblock%2Fqueue-sysfs.txt&data=04%7C01%7Ckrajaram%40geocomputing.net%7C52cc6360e6ea4be737ba08d9f7317d78%7C229a2792a5064f25b3bdbab585cec3ed%7C0%7C0%7C637812615096678246%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C2000&sdata=II8k%2FHzrU7BC%2FVejg9AujgZGk1E0XTz8QCpH6IE6RGM%3D&reserved=0>). But what puzzles me even more: one of the server compiles IOs even smaller, varying between 3.2MiB and 3.6MiB mostly - both for reads and writes ... I just cannot see why. I have to suspect that this will (in writing to the storage) cause incomplete stripe writes on our erasure-coded volumes (8+2p)(as long as the controller is not able to re-coalesce the data properly; and it seems it cannot do it completely at least) If someone of you has seen that already and/or knows a potential explanation I'd be glad to learn about. And if some of you wonder: yes, I (was) moved away from IBM and am now at KIT. Many thanks in advance Uwe-- Karlsruhe Institute of Technology (KIT)Steinbuch Centre for Computing (SCC) Scientific Data Management (SDM) Uwe Falke Hermann-von-Helmholtz-Platz 1, Building 442, Room 187 D-76344 Eggenstein-Leopoldshafen Tel: +49 721 608 28024 Email: uwe.fa...@kit.edu www.scc.kit.edu <https://nam02.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww.scc.kit.edu%2F&data=04%7C01%7Ckrajaram%40geocomputing.net%7C52cc6360e6ea4be737ba08d9f7317d78%7C229a2792a5064f25b3bdbab585cec3ed%7C0%7C0%7C637812615096678246%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C2000&sdata=mXwzkLB1EFB1Dh31rVRMwJZBY4CBbHcJduc9gK6M71A%3D&reserved=0> Registered office: Kaiserstraße 12, 76131 Karlsruhe, Germany KIT – The Research University in the Helmholtz Association _______________________________________________ gpfsug-discuss mailing list gpfsug-discuss at spectrumscale.org <https://nam02.safelinks.protection.outlook.com/?url=http%3A%2F%2Fspectrumscale.org%2F&data=04%7C01%7Ckrajaram%40geocomputing.net%7C52cc6360e6ea4be737ba08d9f7317d78%7C229a2792a5064f25b3bdbab585cec3ed%7C0%7C0%7C637812615096678246%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C2000&sdata=G6bjUWlzkKzR2ptGcLffwD8qF2IT9vkruoevFoTwNE0%3D&reserved=0> http://gpfsug.org/mailman/listinfo/gpfsug-discuss <https://nam02.safelinks.protection.outlook.com/?url=http%3A%2F%2Fgpfsug.org%2Fmailman%2Flistinfo%2Fgpfsug-discuss&data=04%7C01%7Ckrajaram%40geocomputing.net%7C52cc6360e6ea4be737ba08d9f7317d78%7C229a2792a5064f25b3bdbab585cec3ed%7C0%7C0%7C637812615096678246%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C2000&sdata=G9l4PMuzdNA%2BwtAtWK%2BApoXxvKn5jZKeP%2FENOVc9xXg%3D&reserved=0> _______________________________________________ gpfsug-discuss mailing list gpfsug-discuss at spectrumscale.org http://gpfsug.org/mailman/listinfo/gpfsug-discuss <https://nam02.safelinks.protection.outlook.com/?url=http%3A%2F%2Fgpfsug.org%2Fmailman%2Flistinfo%2Fgpfsug-discuss&data=04%7C01%7Ckrajaram%40geocomputing.net%7C52cc6360e6ea4be737ba08d9f7317d78%7C229a2792a5064f25b3bdbab585cec3ed%7C0%7C0%7C637812615096678246%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C2000&sdata=G9l4PMuzdNA%2BwtAtWK%2BApoXxvKn5jZKeP%2FENOVc9xXg%3D&reserved=0> -- Karlsruhe Institute of Technology (KIT) Steinbuch Centre for Computing (SCC) Scientific Data Management (SDM) Uwe Falke Hermann-von-Helmholtz-Platz 1, Building 442, Room 187 D-76344 Eggenstein-Leopoldshafen Tel: +49 721 608 28024 Email:uwe.fa...@kit.edu www.scc.kit.edu <https://nam02.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww.scc.kit.edu%2F&data=04%7C01%7Ckrajaram%40geocomputing.net%7C52cc6360e6ea4be737ba08d9f7317d78%7C229a2792a5064f25b3bdbab585cec3ed%7C0%7C0%7C637812615096678246%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C2000&sdata=mXwzkLB1EFB1Dh31rVRMwJZBY4CBbHcJduc9gK6M71A%3D&reserved=0> Registered office: Kaiserstraße 12, 76131 Karlsruhe, Germany KIT – The Research University in the Helmholtz Association _______________________________________________ gpfsug-discuss mailing list gpfsug-discuss at spectrumscale.org http://gpfsug.org/mailman/listinfo/gpfsug-discuss
-- Karlsruhe Institute of Technology (KIT) Steinbuch Centre for Computing (SCC) Scientific Data Management (SDM) Uwe Falke Hermann-von-Helmholtz-Platz 1, Building 442, Room 187 D-76344 Eggenstein-Leopoldshafen Tel: +49 721 608 28024 Email:uwe.fa...@kit.edu www.scc.kit.edu Registered office: Kaiserstraße 12, 76131 Karlsruhe, Germany KIT – The Research University in the Helmholtz Association
smime.p7s
Description: S/MIME Cryptographic Signature
_______________________________________________ gpfsug-discuss mailing list gpfsug-discuss at spectrumscale.org http://gpfsug.org/mailman/listinfo/gpfsug-discuss
