[lustre-discuss] Lustre Client Lockup Under Buffered I/O (2.14/2.15)

Ellis Wilson elliswilson at microsoft.com
Mon Feb 14 14:27:24 PST 2022 

I believe I have root caused this, and posted detailed analysis on the opened JIRA issue (link in the previous message).  Questions for the community:

1. The Lustre manual claims that "By default, sync_journal is enabled (sync_journal=1), so that journal entries are committed synchronously," but I'm finding that the reverse is and has been true for over a decade.  This is the cause of my client OOM malaise - my clients are holding onto referenced pages until the OSTs commit their journals *and* the clients ping the MGS or somebody else that updates their last committed transaction number to a value greater than the outstanding requests.  These small clients (in fact, even ones as large as 64GB) can easily write fast enough to exhaust memory before the OSTs decide it's time to flush the transactions.  Can somebody clarify if this is just a clerical error in the manual and async journal committing is expected to be default and safe?

2. It appears that although the default "commit" mount option for ext4 is 5 seconds, this is either disabled entirely or set to a much higher value in ldiskfs.  Can somebody clarify what the ldiskfs default setting is for commit (I'm failing hard trying to locate it in code or ldiskfs patches)?  Adjusting the mount option on the OST to use "commit=5" does the right thing (prevents my client from going OOM without the workaround in #1) from what I can tell, so 5s must not be the default for ldiskfs.

3. Are there thoughts from the community on whether setting "sync_journal=1" in lctl or changing the mount option to "commit=5" are preferable?  The latter seems like it will be slightly more performant for very busy systems, but for streaming I/O so far they produce identical results.

4. OFD targets appear to maintain grant info relating to dirty, pending, and current available grant.  I'm witnessing pending well exceed the ldiskfs journal size on my OSTs (which defaults to 1GB).  Code suggests these two are discrete concepts, as pending is correctness checked against blocks in the filesystem shifted left by the power of two associated with the block size.  What's the rationale behind the pending value?

Best,

ellis

From: Ellis Wilson
Sent: Thursday, January 20, 2022 2:28 PM
To: Peter Jones <pjones at whamcloud.com>; Raj <rajgautam at gmail.com>; Patrick Farrell <pfarrell at ddn.com>
Cc: lustre-discuss at lists.lustre.org
Subject: RE: [lustre-discuss] [EXTERNAL] Re: Lustre Client Lockup Under Buffered I/O (2.14/2.15)

Thanks for facilitating a login for me Peter.  The bug with all logs and info I could think to include has been opened here:

https://jira.whamcloud.com/browse/LU-15468

I'm going to keep digging on my end, but if anybody has any other bright ideas or experiments they'd like me to try, don't hesitate to say so here or in the bug.

From: Peter Jones <pjones at whamcloud.com<mailto:pjones at whamcloud.com>>
Sent: Thursday, January 20, 2022 9:28 AM
To: Ellis Wilson <elliswilson at microsoft.com<mailto:elliswilson at microsoft.com>>; Raj <rajgautam at gmail.com<mailto:rajgautam at gmail.com>>; Patrick Farrell <pfarrell at ddn.com<mailto:pfarrell at ddn.com>>
Cc: lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>
Subject: Re: [lustre-discuss] [EXTERNAL] Re: Lustre Client Lockup Under Buffered I/O (2.14/2.15)

You don't often get email from pjones at whamcloud.com<mailto:pjones at whamcloud.com>. Learn why this is important<http://aka.ms/LearnAboutSenderIdentification>
Ellis

JIRA accounts can be requested from info at whamcloud.com<mailto:info at whamcloud.com>

Peter

From: lustre-discuss <lustre-discuss-bounces at lists.lustre.org<mailto:lustre-discuss-bounces at lists.lustre.org>> on behalf of Ellis Wilson via lustre-discuss <lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>>
Reply-To: Ellis Wilson <elliswilson at microsoft.com<mailto:elliswilson at microsoft.com>>
Date: Thursday, January 20, 2022 at 6:20 AM
To: Raj <rajgautam at gmail.com<mailto:rajgautam at gmail.com>>, Patrick Farrell <pfarrell at ddn.com<mailto:pfarrell at ddn.com>>
Cc: "lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>" <lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>>
Subject: Re: [lustre-discuss] [EXTERNAL] Re: Lustre Client Lockup Under Buffered I/O (2.14/2.15)

Thanks Raj - I've checked all of the nodes in the cluster and they all have peer_credits set to 8, and credits are set to 256.  AFAIK that's quite low - 8 concurrent sends to any given peer at a time. Since I only have two OSSes, for this client, that's only 16 concurrent sends at a given moment.  IDK if at this level this devolves to the maximum RPC size of 1MB or the current max BRW I have set of 4MB, but in either case these are small MB values.

I've reached out to Andreas and Patrick to try to get a JIRA account to open a bug, but have not heard back yet.  If somebody on-list is more appropriate to assist with this, please ping me.  I collected quite a bit of logs/traces yesterday and have sysrq stacks to share when I can get access to the whamcloud JIRA.

Best,

ellis

From: Raj <rajgautam at gmail.com<mailto:rajgautam at gmail.com>>
Sent: Thursday, January 20, 2022 8:14 AM
To: Patrick Farrell <pfarrell at ddn.com<mailto:pfarrell at ddn.com>>
Cc: Andreas Dilger <adilger at whamcloud.com<mailto:adilger at whamcloud.com>>; Ellis Wilson <elliswilson at microsoft.com<mailto:elliswilson at microsoft.com>>; lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>
Subject: [EXTERNAL] Re: [lustre-discuss] Lustre Client Lockup Under Buffered I/O (2.14/2.15)

You don't often get email from rajgautam at gmail.com<mailto:rajgautam at gmail.com>. Learn why this is important<http://aka.ms/LearnAboutSenderIdentification>
Ellis, I would also check the peer_credit between server and the client. They should be same.

On Wed, Jan 19, 2022 at 9:27 AM Patrick Farrell via lustre-discuss <lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>> wrote:
Ellis,

As you may have guessed, that function just set looks like a node which is doing buffered I/O and thrashing for memory.  No particular insight available from the count of functions there.

Would you consider opening a bug report in the Whamcloud JIRA?  You should have enough for a good report, here's a few things that would be helpful as well:

It sounds like you can hang the node on demand.  If you could collect stack traces with:

echo t > /proc/sysrq-trigger
after creating the hang, that would be useful.  (It will print to dmesg.)

You've also collected debug logs - Could you include, say, the last 100 MiB of that log set?  That should be reasonable to attach if compressed.

Regards,
Patrick
________________________________
From: lustre-discuss <lustre-discuss-bounces at lists.lustre.org<mailto:lustre-discuss-bounces at lists.lustre.org>> on behalf of Ellis Wilson via lustre-discuss <lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>>
Sent: Wednesday, January 19, 2022 8:32 AM
To: Andreas Dilger <adilger at whamcloud.com<mailto:adilger at whamcloud.com>>
Cc: lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org> <lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>>
Subject: Re: [lustre-discuss] Lustre Client Lockup Under Buffered I/O (2.14/2.15)


Hi Andreas,



Apologies in advance for the top-post.  I'm required to use Outlook for work, and it doesn't handle in-line or bottom-posting well.



Client-side defaults prior to any tuning of mine (this is a very minimal 1-client, 1-MDS/MGS, 2-OSS cluster):

~# lctl get_param llite.*.max_cached_mb

llite.lustrefs-ffff8d52a9c52800.max_cached_mb=

users: 5

max_cached_mb: 7748

used_mb: 0

unused_mb: 7748

reclaim_count: 0

~# lctl get_param osc.*.max_dirty_mb

osc.lustrefs-OST0000-osc-ffff8d52a9c52800.max_dirty_mb=1938

osc.lustrefs-OST0001-osc-ffff8d52a9c52800.max_dirty_mb=1938

~# lctl get_param osc.*.max_rpcs_in_flight

osc.lustrefs-OST0000-osc-ffff8d52a9c52800.max_rpcs_in_flight=8

osc.lustrefs-OST0001-osc-ffff8d52a9c52800.max_rpcs_in_flight=8

~# lctl get_param osc.*.max_pages_per_rpc

osc.lustrefs-OST0000-osc-ffff8d52a9c52800.max_pages_per_rpc=1024

osc.lustrefs-OST0001-osc-ffff8d52a9c52800.max_pages_per_rpc=1024



Thus far I've reduced the following to what I felt were really conservative values for a 16GB RAM machine:



~# lctl set_param llite.*.max_cached_mb=1024

llite.lustrefs-ffff8d52a9c52800.max_cached_mb=1024

~# lctl set_param osc.*.max_dirty_mb=512

osc.lustrefs-OST0000-osc-ffff8d52a9c52800.max_dirty_mb=512

osc.lustrefs-OST0001-osc-ffff8d52a9c52800.max_dirty_mb=512

~# lctl set_param osc.*.max_pages_per_rpc=128

osc.lustrefs-OST0000-osc-ffff8d52a9c52800.max_pages_per_rpc=128

osc.lustrefs-OST0001-osc-ffff8d52a9c52800.max_pages_per_rpc=128

~# lctl set_param osc.*.max_rpcs_in_flight=2

osc.lustrefs-OST0000-osc-ffff8d52a9c52800.max_rpcs_in_flight=2

osc.lustrefs-OST0001-osc-ffff8d52a9c52800.max_rpcs_in_flight=2



This slows down how fast I get to basically OOM from <10 seconds to more like 25 seconds, but the trend is identical.



As an example of what I'm seeing on the client, you can see below we start with most free, and then iozone rapidly (within ~10 seconds) causes all memory to be marked used, and that stabilizes at about 140MB free until at some point it stalls for 20 or more seconds and then some has been synced out:

~# dstat --mem

------memory-usage-----

used  free  buff  cach

1029M 13.9G 2756k  215M

1028M 13.9G 2756k  215M

1028M 13.9G 2756k  215M

1088M 13.9G 2756k  215M

2550M 11.5G 2764k 1238M

3989M 10.1G 2764k 1236M

5404M 8881M 2764k 1239M

6831M 7453M 2772k 1240M

8254M 6033M 2772k 1237M

9672M 4613M 2772k 1239M

10.6G 3462M 2772k 1240M

12.1G 1902M 2772k 1240M

13.4G  582M 2772k 1240M

13.9G  139M 2488k 1161M

13.9G  139M 1528k 1174M

13.9G  140M  896k 1175M

13.9G  139M  676k 1176M

13.9G  142M  528k 1177M

13.9G  140M  484k 1188M

13.9G  139M  492k 1188M

13.9G  139M  488k 1188M

13.9G  141M  488k 1186M

13.9G  141M  480k 1187M

13.9G  139M  492k 1188M

13.9G  141M  600k 1188M

13.9G  139M  580k 1187M

13.9G  140M  536k 1186M

13.9G  141M  668k 1186M

13.9G  139M  580k 1188M

13.9G  140M  568k 1187M

12.7G 1299M 2064k 1197M missed 20 ticks <-- client is totally unresponsive during this time

11.0G 2972M 5404k 1238M^C



Additionally, I've messed with sysctl settings.  Defaults:

vm.dirty_background_bytes = 0

vm.dirty_background_ratio = 10

vm.dirty_bytes = 0

vm.dirty_expire_centisecs = 3000

vm.dirty_ratio = 20

vm.dirty_writeback_centisecs = 500



Revised to conservative values:

vm.dirty_background_bytes = 1073741824

vm.dirty_background_ratio = 0

vm.dirty_bytes = 2147483648

vm.dirty_expire_centisecs = 200

vm.dirty_ratio = 0

vm.dirty_writeback_centisecs = 500



No observed improvement.



I'm going to trawl two logs today side-by-side, one with ldiskfs backing the OSTs, and one with zfs backing the OSTs, and see if I can see what the differences are since the zfs-backed version never gave us this problem.  The only other potentially useful thing I can share right now is that when I turned on full debug logging and ran the test until I hit OOM, the following were the most frequently hit functions in the logs (count, descending, is the first column).  This was approximately 30s of logs:

 205874 cl_page.c:518:cl_vmpage_page())

 206587 cl_page.c:545:cl_page_owner_clear())

 206673 cl_page.c:551:cl_page_owner_clear())

 206748 osc_cache.c:2483:osc_teardown_async_page())

 206815 cl_page.c:867:cl_page_delete())

 206862 cl_page.c:837:cl_page_delete0())

 206878 osc_cache.c:2478:osc_teardown_async_page())

 206928 cl_page.c:869:cl_page_delete())

 206930 cl_page.c:441:cl_page_state_set0())

 206988 osc_page.c:206:osc_page_delete())

 207021 cl_page.c:179:__cl_page_free())

 207021 cl_page.c:193:cl_page_free())

 207021 cl_page.c:532:cl_vmpage_page())

 207024 cl_page.c:210:cl_page_free())

 207075 cl_page.c:430:cl_page_state_set0())

 207169 osc_cache.c:2505:osc_teardown_async_page())

 207175 cl_page.c:475:cl_pagevec_put())

 207202 cl_page.c:492:cl_pagevec_put())

 207211 cl_page.c:822:cl_page_delete0())

 207384 osc_page.c:178:osc_page_delete())

 207422 osc_page.c:177:osc_page_delete())

 413680 cl_page.c:433:cl_page_state_set0())

 413701 cl_page.c:477:cl_pagevec_put())



If anybody has any additional suggestions or requests for more info don't hesitate to ask.



Best,



ellis



From: Andreas Dilger <adilger at whamcloud.com<mailto:adilger at whamcloud.com>>
Sent: Tuesday, January 18, 2022 9:54 PM
To: Ellis Wilson <elliswilson at microsoft.com<mailto:elliswilson at microsoft.com>>
Cc: lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>
Subject: [EXTERNAL] Re: [lustre-discuss] Lustre Client Lockup Under Buffered I/O (2.14/2.15)



You don't often get email from adilger at whamcloud.com<mailto:adilger at whamcloud.com>. Learn why this is important<http://aka.ms/LearnAboutSenderIdentification>

On Jan 18, 2022, at 13:40, Ellis Wilson via lustre-discuss <lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>> wrote:



Recently we've switched from using ZFS to ldiskfs as the backing filesystem to work around some performance issues and I'm finding that when I put the cluster under load (with as little as a single client) I can almost completely lockup the client.  SSH (even existing sessions) stall, iostat, top, etc all freeze for 20 to 200 seconds.  This alleviates for small windows and recurs as long as I leave the io-generating process in existence.  It reports extremely high CPU and RAM usage, and appears to be consumed exclusively doing 'system'-tagged work.  This is on 2.14.0, but I've reproduced on more or less HOL for master-next.  If I do direct-IO, performance is fantastic and I have no such issues regarding CPU/memory pressure.

Uname: Linux 85df894e-8458-4aa4-b16f-1d47154c0dd2-lclient-a0-g0-vm 5.4.0-1065-azure #68~18.04.1-Ubuntu SMP Fri Dec 3 14:08:44 UTC 2021 x86_64 x86_64 x86_64 GNU/Linux

I dmesg I see consistent spew on the client about:
[19548.601651] LustreError: 30918:0:(events.c:208:client_bulk_callback()) event type 1, status -5, desc 00000000b69b83b0
[19548.662647] LustreError: 30917:0:(events.c:208:client_bulk_callback()) event type 1, status -5, desc 000000009ef2fc22
[19549.153590] Lustre: lustrefs-OST0000-osc-ffff8d52a9c52800: Connection to lustrefs-OST0000 (at 10.1.98.7 at tcp<mailto:10.1.98.7 at tcp>) was lost; in progress operations using this service will wait for recovery to complete
[19549.153621] Lustre: 30927:0:(client.c:2282:ptlrpc_expire_one_request()) @@@ Request sent has failed due to network error: [sent 1642535831/real 1642535833]  req at 0000000002361e2d x1722317313374336/t0(0) o4->lustrefs-OST0001-osc-ffff8d52a9c52800 at 10.1.98.10<mailto:lustrefs-OST0001-osc-ffff8d52a9c52800 at 10.1.98.10>@tcp:6/4 lens 488/448 e 0 to 1 dl 1642535883 ref 2 fl Rpc:eXQr/0/ffffffff rc 0/-1 job:''
[19549.153623] Lustre: 30927:0:(client.c:2282:ptlrpc_expire_one_request()) Skipped 4 previous similar messages

But I actually think this is a symptom of extreme memory pressure causing the client to timeout things, not a cause.

Testing with obdfilter-survey (local) on the OSS side shows expected performance of the disk subsystem.  Testing with lnet_selftest from client to OSS shows expected performance.  In neither case do I see the high cpu or memory pressure issues.

Reducing a variety of lctl tunables that appear to govern memory allowances for Lustre clients does not improve the situation.



What have you reduced here?  llite.*.max_cached_mb, osc.*.max_dirty_mb, osc.*.max_rpcs_in_flight and osc.*.max_pages_per_rpc?



By all appearances, the running iozone or even simple dd processes gradually (i.e., over a span of just 10 seconds or so) consumes all 16GB of RAM on the client I'm using.  I've generated bcc profile graphs for both on- and off-cpu analysis, and they are utterly boring -- they basically just reflect rampant calls to shrink_inactive_list resulting from page_cache_alloc in the presence of extreme memory pressure.



We have seen some issues like this that are being looked at, but this is mostly only seen on smaller VM clients used in testing and not larger production clients.  Are you able to test with more RAM on the client?  Have you tried with 2.12.8 installed on the client?



Cheers, Andreas

--

Andreas Dilger

Lustre Principal Architect

Whamcloud












_______________________________________________
lustre-discuss mailing list
lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>
http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org<https://nam06.safelinks.protection.outlook.com/?url=http%3A%2F%2Flists.lustre.org%2Flistinfo.cgi%2Flustre-discuss-lustre.org&data=04%7C01%7Celliswilson%40microsoft.com%7Cfa5278382e5642deae0208d9dc210892%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637782856719975502%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=DT%2Bb%2BQ1ec7rQcLhU1Pm9p60JHNQTZKQq51hRT2zouLc%3D&reserved=0>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/attachments/20220214/1eddcaf1/attachment-0001.html>

More information about the lustre-discuss mailing list