8

We have several redis instances running on a server. There are also multiple web tier servers connecting to those instance that experience a stall at the same time.

We had packet captures going at the time, that identified that there wall a stall in both TX and RX traffic, as per the following wireshark IO graphs:

enter image description here

enter image description here

There was a correlating spike in redis calls, but I suspect that was an effect and not a cause due to the time lag:

enter image description here

With a 15/s sampling interval (this is collected as a counter) there was an average of 136 memory allocation stalls:

enter image description here

There also seemed to be an out of the ordinary number of NUMA pages migrated at that same time:

enter image description here

Although the above looks normal, there were two consecutive data points for this which make it abnormal compared to other above 300 spikes seen in the graph.

There was also correlated spike in memory compaction failures and compaction stalls:

enter image description here

enter image description here

Although I have a wealth of memory information here, my Linux memory knowledge is not deep enough to really hypothesize a good story that brings all this information together to explain the stall. Can anyone with deep Linux memory knowledge (and perhaps deep redis memory knowledge) tie some of this information together?

We collect all stats from /proc/vmstat at 15 second intervals, so if there is any data there that you think might add to this please do request it. I just picked the things that seemed to have interesting activity, in particular the alloc stall, the numa migration, and the compaction stall/failures. The totals follow, and cover 20 days of uptime:

[kbrandt@ny-redis01: ~] uptime
 21:11:49 up 20 days, 20:05,  8 users,  load average: 1.05, 0.74, 0.69
[kbrandt@ny-redis01: ~] cat /proc/vmstat
nr_free_pages 105382
nr_alloc_batch 5632
nr_inactive_anon 983455
nr_active_anon 15870487
nr_inactive_file 12904618
nr_active_file 2266184
nr_unevictable 0
nr_mlock 0
nr_anon_pages 16361259
nr_mapped 26329
nr_file_pages 15667318
nr_dirty 48588
nr_writeback 0
nr_slab_reclaimable 473720
nr_slab_unreclaimable 37147
nr_page_table_pages 38701
nr_kernel_stack 987
nr_unstable 0
nr_bounce 0
nr_vmscan_write 356302
nr_vmscan_immediate_reclaim 174305
nr_writeback_temp 0
nr_isolated_anon 0
nr_isolated_file 32
nr_shmem 423906
nr_dirtied 3071978326
nr_written 3069010459
numa_hit 1825289996
numa_miss 3360625955
numa_foreign 3360626253
numa_interleave 64798
numa_local 1856473774
numa_other 3329442177
workingset_refault 297175
workingset_activate 24923
workingset_nodereclaim 0
nr_anon_transparent_hugepages 41
nr_free_cma 0
nr_dirty_threshold 3030688
nr_dirty_background_threshold 1515344
pgpgin 25709012
pgpgout 12284206511
pswpin 143954
pswpout 341570
pgalloc_dma 430
pgalloc_dma32 498407404
pgalloc_normal 8131576449
pgalloc_movable 0
pgfree 8639210186
pgactivate 12022290
pgdeactivate 14512106
pgfault 61444049878
pgmajfault 23740
pgrefill_dma 0
pgrefill_dma32 1084722
pgrefill_normal 13419119
pgrefill_movable 0
pgsteal_kswapd_dma 0
pgsteal_kswapd_dma32 11991303
pgsteal_kswapd_normal 1051781383
pgsteal_kswapd_movable 0
pgsteal_direct_dma 0
pgsteal_direct_dma32 58737
pgsteal_direct_normal 36277968
pgsteal_direct_movable 0
pgscan_kswapd_dma 0
pgscan_kswapd_dma32 13416911
pgscan_kswapd_normal 1053143529
pgscan_kswapd_movable 0
pgscan_direct_dma 0
pgscan_direct_dma32 58926
pgscan_direct_normal 36291030
pgscan_direct_movable 0
pgscan_direct_throttle 0
zone_reclaim_failed 0
pginodesteal 0
slabs_scanned 1812992
kswapd_inodesteal 5096998
kswapd_low_wmark_hit_quickly 8600243
kswapd_high_wmark_hit_quickly 5068337
pageoutrun 14095945
allocstall 567491
pgrotated 971171
drop_pagecache 8
drop_slab 0
numa_pte_updates 58218081649
numa_huge_pte_updates 416664
numa_hint_faults 57988385456
numa_hint_faults_local 57286615202
numa_pages_migrated 39923112
pgmigrate_success 48662606
pgmigrate_fail 2670596
compact_migrate_scanned 29140124
compact_free_scanned 28320190101
compact_isolated 21473591
compact_stall 57784
compact_fail 37819
compact_success 19965
htlb_buddy_alloc_success 0
htlb_buddy_alloc_fail 0
unevictable_pgs_culled 5528
unevictable_pgs_scanned 0
unevictable_pgs_rescued 18567
unevictable_pgs_mlocked 20909
unevictable_pgs_munlocked 20909
unevictable_pgs_cleared 0
unevictable_pgs_stranded 0
thp_fault_alloc 11613
thp_fault_fallback 53
thp_collapse_alloc 3
thp_collapse_alloc_failed 0
thp_split 9804
thp_zero_page_alloc 1
thp_zero_page_alloc_failed 0 

Also all /proc/sys/vm/* settings if that helps:

***/proc/sys/vm/admin_reserve_kbytes***
8192
***/proc/sys/vm/block_dump***
0
***/proc/sys/vm/dirty_background_bytes***
0
***/proc/sys/vm/dirty_background_ratio***
10
***/proc/sys/vm/dirty_bytes***
0
***/proc/sys/vm/dirty_expire_centisecs***
3000
***/proc/sys/vm/dirty_ratio***
20
***/proc/sys/vm/dirty_writeback_centisecs***
500
***/proc/sys/vm/drop_caches***
1
***/proc/sys/vm/extfrag_threshold***
500
***/proc/sys/vm/hugepages_treat_as_movable***
0
***/proc/sys/vm/hugetlb_shm_group***
0
***/proc/sys/vm/laptop_mode***
0
***/proc/sys/vm/legacy_va_layout***
0
***/proc/sys/vm/lowmem_reserve_ratio***
256 256 32
***/proc/sys/vm/max_map_count***
65530
***/proc/sys/vm/memory_failure_early_kill***
0
***/proc/sys/vm/memory_failure_recovery***
1
***/proc/sys/vm/min_free_kbytes***
90112
***/proc/sys/vm/min_slab_ratio***
5
***/proc/sys/vm/min_unmapped_ratio***
1
***/proc/sys/vm/mmap_min_addr***
4096
***/proc/sys/vm/nr_hugepages***
0
***/proc/sys/vm/nr_hugepages_mempolicy***
0
***/proc/sys/vm/nr_overcommit_hugepages***
0
***/proc/sys/vm/nr_pdflush_threads***
0
***/proc/sys/vm/numa_zonelist_order***
default
***/proc/sys/vm/oom_dump_tasks***
1
***/proc/sys/vm/oom_kill_allocating_task***
0
***/proc/sys/vm/overcommit_kbytes***
0
***/proc/sys/vm/overcommit_memory***
1
***/proc/sys/vm/overcommit_ratio***
50
***/proc/sys/vm/page-cluster***
3
***/proc/sys/vm/panic_on_oom***
0
***/proc/sys/vm/percpu_pagelist_fraction***
0
***/proc/sys/vm/scan_unevictable_pages***
0
***/proc/sys/vm/stat_interval***
1
***/proc/sys/vm/swappiness***
60
***/proc/sys/vm/user_reserve_kbytes***
131072
***/proc/sys/vm/vfs_cache_pressure***
100
***/proc/sys/vm/zone_reclaim_mode***
0

Update:

There is a thp_split that is close in time:

enter image description here

2 Answers 2

4

What is your setting for /proc/sys/vm/zone_reclaim? Try setting it to 0. There's plenty of stuff on the net if you search for 'zone_reclaim', so I won't try to rehash it here.

5
  • Is already zero it seems [kbrandt@ny-redis01: ~] cat /proc/sys/vm/zone_reclaim_mode 0 Jul 28, 2015 at 21:38
  • 1
    OK. Next stop: transparent hugepages. Looks like they're in use: thp_fault_alloc 11613 thp_fault_fallback 53 thp_collapse_alloc 3 thp_collapse_alloc_failed 0 thp_split 9804 thp_zero_page_alloc 1 thp_zero_page_alloc_failed 0 For Redis you probably want THP disabled. antirez.com/news/84 TL;DR: echo never > /sys/kernel/mm/transparent_hugepage/enabled
    – Tobert
    Jul 28, 2015 at 21:49
  • There is a thp_split that is a close correlation. Some lag in recording but the monitoring agent timing could have been impacted as well. Updated the question with the graph! Jul 29, 2015 at 0:35
  • Mind enlightening us to what is going on during a thp split? Jul 29, 2015 at 0:37
  • We have it set in puppet, but as an exec. I don't see it in rc.local, so my current hypothesis is a race condition - redis starts before the kernel switch is set Jul 29, 2015 at 1:44
2

When Redis forks to checkpoint, the Linux kernel needs to duplicate the mapping tables for copy on write. If you have a lot of RAM, this can take a lot of time. We have a 200 GB Redis instance that takes 8 seconds to fork, and the machine is deaf to the world while this happens.

Workarounds (from easy to hard):

  • checkpoint less often, increasing the time and key count before checkpoint
  • shard your data into multiple process instances, each of which uses less RAM
  • try aof instead of checkpoint, although this will fork occasionally anyway
  • try huge pages, although you may need to double your physical RAM because approximately everything will be dirtied while checkpointing
  • screw it and go with Postgres

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .