https://jira.whamcloud.com This file is an XML representation of an issue en-us 9.4.14 940014 05-12-2023 https://jira.whamcloud.com/browse/LU-7293 Lustre <p>I did a detail study of the client scaling behavior for 10k and 100k files per directory using 1,2,4, and 8 MDS servers each having one MDT. I also attempted to collect dat for 16 MDS servers but the results were so bad I didn't both to finish collecting them since it would take several months to finish the 16 node case.</p> DNE2 system with up to 16 MDS servers. Uses up to 400 client nodes spread across 20 physical nodes. All the results are based on mdtest 1.9.3 runs. LU-7293

DNE2 perfomance analysis

Improvement Critical Resolved Fixed Di Wang James A Simmons dne2 dne3 Tue, 13 Oct 2015 22:45:13 +0000 Sat, 12 Jan 2019 04:02:24 +0000 Sat, 12 Jan 2019 04:02:24 +0000 Lustre 2.9.0 0 14 <p>James, Could you please tell me more of your test? commands? "MDS striping 8" means the directory stripe_count = 8 ? Just curious, you did not disable quota in your test right? Thanks</p> <p>I attached my job script I used. You just need to replace apron with mpirun with the correct node count and thread count. As for quotas I haven't touch that setting. In the past it was off by default. Is this not the case anymore? Yes the MDS striping is the lfs setdirstripe -c value I used.</p> <p>I think quota is enable by default since 2.4. To disable quota you have to "tune2fs -O ^quota" after reformat. But since there are only 1 MDT per MDS, quota is probably irrelevant. Thanks I will check the script. </p> <p>In mdtest-scale.pbs</p> <div class="preformatted panel" style="border-width: 1px;"><div class="preformattedContent panelContent"> <pre>!/bin/bash #PBS -l nodes=20 #PBS -l walltime=24:00:00 #PBS -N results-mdtest-scale #PBS -j oe MOUNT=sultan OSTCOUNT=$(lctl get_param -n lov.$MOUNT-clilov*.numobd) ITER=5 PBS_JOBID="dne2_8_mds" BINDIR=/lustre/$MOUNT/stf008/scratch/$USER OUTDIR=$BINDIR/${PBS_JOBID}_md_test [ -e $OUTDIR ] || { mkdir -p $OUTDIR lfs setstripe -c $OSTCOUNT $OUTDIR } cd $BINDIR </pre> </div></div> <p>It seems this script is for testing OST stripes? you probably post the wrong script? thanks</p> <p>The script always sets the test directories so all created files are striped across all the OSTs. I do this with all my tests so when I move to our large stripe test setup I can test to ensure 1008 OST stripe files work as well. It also ensure that my fail over testing will always cover active servers on the OSS side. </p> <p>oh, I want to know how the directories are striped among MDT? Did you set default dir stripes (or /lustre/$MOUNT/stf008/scratch/$USER ? If you did, then it probably explain why directory creation is much slower for multiple MDT, because all of directories created here are striped directory. Anyway, could you please tell me how BINDIR/OUTDIR are striped here? Thanks.</p> <p>I precreated each striped/split directory before each run of the mdtest-scale.pbs script. Each directory was called dne2_"stripe_count"_mds. I set remote_dir=1 and remote_dir_gid=-1 so the normal user (as myself) could create the striped directory.<br/> I did the follow before each test set using 4 as a example:</p> <p>lfs setdirstripe -c 4 --index=4 /lustre/sultan/stf008/scratch/jsimmons/dne2_4_mds<br/> lfs setdirstripe -c 4 --index=4 -D /lustre/sultan/stf008/scratch/jsimmons/dne2_4_mds</p> <p>Then I ran the test. I used the -D options so the directories created by mdtest would be the same as dne2_X_mds. I also used the index to avoid filling up my MDS disk with inodes.</p> <p>James: Could you please re-run the test with -F, so only do file test. I want to see if it can show liner performance improvement with only file operation. I suspect there are some interfere between cross-MDT and single MDT operation. </p> <p>And also any reason you use "--index=4 -D" for default stripedEA? That means all of dir create request will be sent MDT4, which might not be what you want? I would suggest remove --index=4 for default striped EA. i.e. <br/> lfs setdirstripe -c4 -D /lustre/sultan/stf008/scratch/jsimmons/dne2_4_mds</p> <p>James, it also doesn't make sense to create huge numbers of regular files striped across all OSTs. Applications should either create large numbers of files with one stripe per file (i.e. file per process) or create small numbers of widely striped files (i.e. shared single file). Doing both at the same time is IMHO not testing what happens on most systems.</p> <p>Having large numbers of widely-striped files both stresses OST object creation rates, as well as slowing down the MDS because it needs to store an extra xattr for each file. In your case with 1008 OSTs, this is actually creating two inodes per file on the MDT in order to store the large xattr (about 24KB/file).</p> <p>This is a test on a small system with only 56 OSTs. I can try the default stripe of 4 but I don't expect that much of difference, Our users tend to do one of two things. Use the default setting or do a lsf setstripe -c -1. Also the goal here was to how scaling behaved.</p> <p>DiWang our test system is under going a upgrade. It will be a few days before it is finished.</p> <p>We are in the process of installing perf on our test systems to analysis what is going on. I should have something next week.</p> <p>I have some good news and some bad news with my testing with perf installed. The good news is I'm seeing much better performance so far with large MDS stripe count. I will start collecting new data soon and post it here. The bad news is when creating one million plus files I'm seeing constant client eviction and reconnects due to time outs from the OSS. I will open a separate ticket for that.</p> <p>Sorry but I was wrong about performance fixes. My script had a error in it where the default stripe was 1 for mdtest. I did do some profiling and I'm not seeing anything hogging cycles. What I did see is this with slab top on my client:</p> <p> Active / Total Objects (% used) : 549437 / 559744 (98.2%)<br/> Active / Total Slabs (% used) : 67767 / 67768 (100.0%)<br/> Active / Total Caches (% used) : 108 / 227 (47.6%)<br/> Active / Total Size (% used) : 532288.63K / 534120.31K (99.7%)<br/> Minimum / Average / Maximum Object : 0.02K / 0.95K / 32768.00K</p> <p>372512 372485 99% 1.12K 53216 7 425728K lustre_inode_cache<br/> 41552 41431 99% 0.03K 371 112 1484K size-32<br/> 38600 38575 99% 0.19K 1930 20 7720K dentry<br/> 15744 15740 99% 0.99K 3936 4 15744K nfs_inode_cache<br/> 14400 14349 99% 0.12K 480 30 1920K size-128<br/> 13098 12493 95% 0.06K 222 59 888K size-64</p> <div class="preformatted panel" style="border-width: 1px;"><div class="preformattedContent panelContent"> <pre>Sorry but I was wrong about performance fixes. </pre> </div></div> <p>So you mean "seeing much better performance" is not correct? will you redo the test? Actually I hope you can use Oprofile to profile MDS, so I can see which function or lock is being hit most, then know where is the bottleneck for this load. Thanks.</p> <p>Here is the perf data I gathered on one of the MDS being used:</p> <p>Samples: 136K of event 'cycles', Event count (approx.): 45672465691<br/> 62.57% ossec-syscheckd ossec-syscheckd <span class="error">&#91;.&#93;</span> 0x0000000000027003<br/> + 3.20% init <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> intel_idle<br/> + 1.57% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> intel_idle<br/> + 0.88% ossec-syscheckd libc-2.12.so <span class="error">&#91;.&#93;</span> memcpy<br/> + 0.78% ossec-syscheckd <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> native_flush_tlb_global<br/> + 0.67% ossec-syscheckd <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> copy_user_generic_string<br/> + 0.61% init <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> poll_idle<br/> + 0.36% init <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> rb_erase<br/> + 0.33% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> _spin_unlock_irqrestore<br/> + 0.24% init <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> _spin_unlock_irqrestore<br/> + 0.22% init <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> __hrtimer_start_range_ns<br/> + 0.22% ossec-syscheckd <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> clear_page_c<br/> + 0.21% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> rb_erase<br/> + 0.21% ossec-syscheckd libc-2.12.so <span class="error">&#91;.&#93;</span> vfprintf<br/> + 0.21% ossec-syscheckd <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> page_fault<br/> + 0.18% ossec-syscheckd <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> _spin_lock<br/> + 0.16% ossec-syscheckd <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> file_read_actor<br/> + 0.15% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> __remove_hrtimer<br/> + 0.15% ossec-syscheckd <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> rpcauth_lookup_credcache<br/> + 0.14% ossec-syscheckd <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> __change_page_attr_set_clr<br/> + 0.14% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> __hrtimer_start_range_ns<br/> + 0.13% init <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> schedule<br/> + 0.13% ossec-syscheckd <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> lookup_address<br/> + 0.13% ossec-syscheckd <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> __alloc_pages_nodemask<br/> + 0.12% init <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> tick_nohz_stop_sched_tick<br/> + 0.11% kiblnd_sd_00_00 <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> find_busiest_group<br/> + 0.11% ossec-syscheckd <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> kmem_cache_alloc<br/> + 0.11% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> poll_idle<br/> + 0.10% init <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> cpuidle_idle_call<br/> + 0.10% ossec-syscheckd <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> free_hot_cold_page</p> <p>and the slab top gives:</p> <p> Active / Total Objects (% used) : 8211619 / 11280489 (72.8%)<br/> Active / Total Slabs (% used) : 511008 / 511020 (100.0%)<br/> Active / Total Caches (% used) : 127 / 279 (45.5%)<br/> Active / Total Size (% used) : 1746332.12K / 2151220.29K (81.2%)<br/> Minimum / Average / Maximum Object : 0.02K / 0.19K / 4096.00K</p> <p> OBJS ACTIVE USE OBJ SIZE SLABS OBJ/SLAB CACHE SIZE NAME <br/> 5973206 5972436 99% 0.10K 161438 37 645752K buffer_head<br/> 1255968 255927 20% 0.12K 39249 32 156996K lod_obj<br/> 1255968 255927 20% 0.08K 26166 48 104664K mdd_obj<br/> 1255843 255927 20% 0.20K 66097 19 264388K mdt_obj<br/> 281643 281643 100% 1.02K 93881 3 375524K ldiskfs_inode_cache<br/> 255892 255853 99% 0.20K 13468 19 53872K osp_obj<br/> 219184 174975 79% 0.03K 1957 112 7828K size-32<br/> 175761 173927 98% 0.06K 2979 59 11916K size-64<br/> 156680 155264 99% 0.19K 7834 20 31336K dentry<br/> 147532 147478 99% 0.55K 21076 7 84304K radix_tree_node<br/> 110262 110257 99% 1.02K 36754 3 147016K nfs_inode_cache<br/> 22914 22876 99% 0.64K 3819 6 15276K proc_inode_cache</p> <p>From this the MDS doesn't look to have too heavy of a burden. Its not cpu pegged nor no memory exhaustion.</p> <p>On the client side I see with perf and slab top:</p> <p>+ 30.58% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> __schedule -<br/> + 12.16% pbs_mom <span class="error">&#91;unknown&#93;</span> <span class="error">&#91;.&#93;</span> 0x7f198dcd6ea3 0<br/> + 6.21% apsys <span class="error">&#91;unknown&#93;</span> <span class="error">&#91;.&#93;</span> 0x407a76 ▒<br/> + 2.03% ptlrpcd_00_02 <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> __switch_to ▒<br/> + 1.32% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> read_tsc ▒<br/> + 1.04% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> load_balance ▒<br/> + 0.93% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> sched_clock_local ▒<br/> + 0.90% kworker/3:1 <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> rt_worker_func ▒<br/> + 0.86% pbs_mom <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> copy_pte_range ▒<br/> + 0.74% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> default_idle ▒<br/> + 0.73% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> run_timer_softirq ▒<br/> + 0.63% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> trace_hardirqs_off ▒<br/> + 0.58% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> ktime_get ▒<br/> + 0.55% pbs_mom <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> vsnprintf ▒<br/> + 0.42% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> get_next_timer_interrupt ▒<br/> + 0.40% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> apic_timer_interrupt ▒<br/> + 0.37% ptlrpcd_00_00 <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> _raw_spin_lock ▒<br/> + 0.36% munged <span class="error">&#91;unknown&#93;</span> <span class="error">&#91;.&#93;</span> 0x409380 ▒<br/> + 0.34% apsys <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> page_fault ▒<br/> + 0.33% pbs_mom <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> number ▒<br/> + 0.33% nscd <span class="error">&#91;unknown&#93;</span> <span class="error">&#91;.&#93;</span> 0x7ffb57c0866f ▒<br/> + 0.32% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> amd_e400_idle ▒<br/> + 0.32% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> _raw_spin_lock ▒<br/> + 0.30% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> cpumask_next_and ▒<br/> + 0.29% pbs_mom <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> __do_fault ▒<br/> + 0.29% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> rcu_exit_nohz ▒<br/> + 0.28% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> x86_pmu_enable_all ▒<br/> + 0.27% swapper <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> native_sched_clock ▒<br/> + 0.26% pbs_mom <span class="error">&#91;kernel.kallsyms&#93;</span> <span class="error">&#91;k&#93;</span> do_task_stat </p> <p> Active / Total Objects (% used) : 6238442 / 6248913 (99.8%)<br/> Active / Total Slabs (% used) : 878492 / 878500 (100.0%)<br/> Active / Total Caches (% used) : 108 / 227 (47.6%)<br/> Active / Total Size (% used) : 6911389.05K / 6913362.17K (100.0%)<br/> Minimum / Average / Maximum Object : 0.02K / 1.11K / 32768.00K</p> <p> OBJS ACTIVE USE OBJ SIZE SLABS OBJ/SLAB CACHE SIZE NAME <br/> 6038739 6038710 99% 1.12K 862677 7 6901416K lustre_inode_cache<br/> 49820 49806 99% 0.19K 2491 20 9964K dentry<br/> 48160 47939 99% 0.03K 430 112 1720K size-32<br/> 18236 18232 99% 0.99K 4559 4 18236K nfs_inode_cache<br/> 17040 16731 98% 0.12K 568 30 2272K size-128<br/> 13806 13346 96% 0.06K 234 59 936K size-64<br/> 7832 7790 99% 0.09K 178 44 712K sysfs_dir_cache<br/> 7329 7309 99% 0.57K 1047 7 4188K inode_cache<br/> 6440 6276 97% 0.19K 322 20 1288K size-192<br/> 5159 2745 53% 0.05K 67 77 268K anon_vma_chain<br/> 3976 3761 94% 0.50K 497 8 1988K size-512</p> <p>Here you can see the client nodes are pegged due to swapper running constantly due to the memory pressure on the client. The lustre_inode_cache is huge.</p> <p>James: did this only happen when you do directory creation? or both file and directory creation? how many threads per client? Thanks</p> <p>Btw: did you rerun the test with "-F" and remove --index for default stripeEA as comment in Oct 18th. Thanks</p> <p>I did both file only operations and with directory operations. I did tracked down the issue of directory operations as well. What is happening in that case is that the lustre inode cache is consuming all the memory on the client thus causing various timeout and client evictions and reconnects. This only happens for when many directory operations are performed. When only doing file operations the memory pressure issues go away. My latest testings as all been without the --index.</p> <p>Hmm, according to the slab information in Nov 15th, it seems "lustre_inode_cache" is much more than "inode_cache", so it means client has more ll_inode_info than inode, hmm, maybe ll_inode_info is leaked somewhere. Do you still keep that client? Could you please get lru_size for me?</p> <div class="preformatted panel" style="border-width: 1px;"><div class="preformattedContent panelContent"> <pre>lctl get_param ldlm.*.*MDT*.lru_size </pre> </div></div> <p>Oh this doesn't look right.</p> <p>ldlm.namespaces.sultan-MDT0000-mdc-ffff8803f3d12c00.lru_size=29<br/> ldlm.namespaces.sultan-MDT0001-mdc-ffff8803f3d12c00.lru_size=0<br/> ldlm.namespaces.sultan-MDT0002-mdc-ffff8803f3d12c00.lru_size=0<br/> ldlm.namespaces.sultan-MDT0003-mdc-ffff8803f3d12c00.lru_size=0<br/> ldlm.namespaces.sultan-MDT0004-mdc-ffff8803f3d12c00.lru_size=0<br/> ldlm.namespaces.sultan-MDT0005-mdc-ffff8803f3d12c00.lru_size=0<br/> ldlm.namespaces.sultan-MDT0006-mdc-ffff8803f3d12c00.lru_size=0<br/> ldlm.namespaces.sultan-MDT0007-mdc-ffff8803f3d12c00.lru_size=0<br/> ldlm.namespaces.sultan-MDT0008-mdc-ffff8803f3d12c00.lru_size=0<br/> ldlm.namespaces.sultan-MDT0009-mdc-ffff8803f3d12c00.lru_size=0<br/> ldlm.namespaces.sultan-MDT000a-mdc-ffff8803f3d12c00.lru_size=0<br/> ldlm.namespaces.sultan-MDT000b-mdc-ffff8803f3d12c00.lru_size=0<br/> ldlm.namespaces.sultan-MDT000c-mdc-ffff8803f3d12c00.lru_size=0<br/> ldlm.namespaces.sultan-MDT000d-mdc-ffff8803f3d12c00.lru_size=0<br/> ldlm.namespaces.sultan-MDT000e-mdc-ffff8803f3d12c00.lru_size=0<br/> ldlm.namespaces.sultan-MDT000f-mdc-ffff8803f3d12c00.lru_size=0</p> <p>Here is the final report of our results for our DNE2 performance analysis</p> <p><a href="http://info.ornl.gov/sites/publications/Files/Pub59510.pdf" class="external-link" target="_blank" rel="nofollow noopener">http://info.ornl.gov/sites/publications/Files/Pub59510.pdf</a></p> <p>Enjoy the read. Perhaps we can link it to the wiki. If people want it linked to the wiki we can do that<br/> and then close this ticket. If not we can keep this ticket open for a few more weeks so people can<br/> have a chance to read this.</p> <p>Hi James,</p> <p>Thank you very much for the report!</p> <p>closing ancient ticket</p> Related LU-6831 Development End date Mon, 21 Dec 2015 22:45:13 +0000 Epic/Theme Performance Rank 1|hzxqe7: Rank (Obsolete) 9223372036854775807 Start date Tue, 13 Oct 2015 22:45:13 +0000