Regression caused by using node_to_bdi()
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
Hi, Christoph
*From: Zhao Lei [mailto:[email protected]]
Quoted 19 lines
Sent: Monday, March 09, 2015 10:47 AM
To: 'Christoph Hellwig'; 'Jan Kara'
Cc: 'Tejun Heo'; 'Jens Axboe'
Subject: RE: Regression caused by usingnode_to_bdi()Hi, Christoph and Jan
* From: 'Christoph Hellwig' [mailto:[email protected]]
Quoted 11 lines
Sent: Sunday, March 08, 2015 11:34 PM
To: Jan Kara
Cc: Zhao Lei; 'Christoph Hellwig'; 'Tejun Heo'; 'Jens Axboe'
Subject: Re: Regression caused by usingnode_to_bdi()Quoted 5 lines · Jan Kara
Frankly, I doubt the cost of
inode_to_bdi()is the reason for the
slowdown here. If I read the numbers right, the throughput dropped from 135 MB/s on average to 130 MB/s on average. Such load is hardly going to saturate the CPU enough for additional cycles ininode_to_bdi()to
Quoted 6 lines
Quoted 4 lines
matter.
Quoted 3 lines
The load like this is completely IO bound unless you have really fast drive (doing GB/s). What are the throughput number just before / after this commit?\
These are performance data before and after this patch In bisect:
What is your opinion about this regression?
Please tell me if you need additional test and result on my env.
Thanks
Zhaolei
Quoted 185 lines
v3.19-rc5_00005_495a27 :
io_speed: valcnt=10 avg=137.409 range=[134.820,139.000] diff=3.10% stdev=1.574 cv=1.15% v3.19-rc5_00006_26ff13 :io_speed: valcnt=10 avg=136.534 range=[132.390,139.500] diff=5.37% stdev=2.659 cv=1.95% v3.19-rc5_00007_de1414 :io_speed: valcnt=10 avg=130.358 range=[129.070,132.150] diff=2.39% stdev=1.120 cv=0.86% <- *this patch* v3.19-rc5_00008_b83ae6 :io_speed: valcnt=10 avg=129.549 range=[129.200,129.910] diff=0.55% stdev=0.241 cv=0.19% v3.19-rc5_00011_c4db59 :io_speed: valcnt=10 avg=130.033 range=[129.050,131.620] diff=1.99% stdev=0.854 cv=0.66%Quoted 2 lines
What is the CPU load while the benchmark is running?
I hadn't record cpu load in testing, I'll do it if it is necessary for debug.
These are of one of sysbench's log:
sysbench 0.4.12: multi-threaded system evaluation benchmark
1 files, 4194304Kb each, 4096Mb total
Creating files for the test...
sysbench 0.4.12: multi-threaded system evaluation benchmarkRunning the test with following options:
Number of threads: 1Extra file open flags: 0
1 files, 4Gb each
4Gb total file size
Block size 32Kb
Using synchronous I/O mode
Doing sequential write (creation) test
Threads started!
Done.Operations performed: 0 Read, 131072 Write, 0 Other = 131072 Total Read 0b Written 4Gb Total transferred 4Gb (132.15Mb/sec)
4228.75 Requests/sec executedTest execution summary:
total time: 30.9955s
total number of events: 131072
total time taken by event execution: 30.8731
per-request statistics:
min: 0.01ms
avg: 0.24ms
max: 30.80ms
approx. 95 percentile: 0.03msThreads fairness:
events (avg/stddev): 131072.0000/0.00 execution time (avg/stddev): 30.8731/0.00sysbench 0.4.12: multi-threaded system evaluation benchmark
Quoted 2 lines
How much memory does the machine have?
2G mem, 2-core machin, test is running on 1T sata disk.
[root@btrfs test_nosync_32768__sync_1_seqwr_4G_btrfs_1]# cat
/proc/meminfoMemTotal: 2015812 kB MemFree: 627416 kB MemAvailable: 1755488 kB Buffers: 345876 kB Cached: 772788 kB SwapCached: 0 kB Active: 848864 kB Inactive: 320044 kB Active(anon): 54128 kB Inactive(anon): 5080 kB Active(file): 794736 kB Inactive(file): 314964 kB Unevictable: 0 kB Mlocked: 0 kB SwapTotal: 0 kB SwapFree: 0 kB Dirty: 0 kB Writeback: 0 kB AnonPages: 50140 kB Mapped: 41636 kB Shmem: 8984 kB Slab: 200312 kB SReclaimable: 187308 kB SUnreclaim: 13004 kB KernelStack: 1728 kB PageTables: 4056 kB NFS_Unstable: 0 kB Bounce: 0 kB WritebackTmp: 0 kB CommitLimit: 1007904 kB Committed_AS: 205956 kB VmallocTotal: 34359738367 kB VmallocUsed: 539968 kB VmallocChunk: 34359195223 kB HardwareCorrupted: 0 kB AnonHugePages: 6144 kB HugePages_Total: 0 HugePages_Free: 0 HugePages_Rsvd: 0 HugePages_Surp: 0 Hugepagesize: 2048 kB DirectMap4k: 61056 kB DirectMap2M: 2000896 kB[root@btrfs test_nosync_32768__sync_1_seqwr_4G_btrfs_1]# cat
/proc/cpuinfo
processor : 0vendor_id: GenuineIntel
cpu family : 6
model : 23
model name : Intel(R) Core(TM)2 Duo CPU E7500 @ 2.93GHz
stepping : 10
microcode :0xa0b
cpu MHz : 1603.000
cache size : 3072 KB
physical id : 0
siblings : 2
core id : 0
cpu cores : 2
apicid : 0
initial apicid : 0
fpu : yesfpu_exception: yes
cpuid level : 13
wp : yes
flags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca
cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx lmconstant_tscarch_perfmonpebs btsrep_goodnopl aperfmperf pni dtes64 monitords_cplvmx est tm2 ssse3 cx16 xtpr pdcmsse4_1xsavelahf_lmdthermtpr_shadowvnmi flexprioritybugs : bogomips : 5851.89 clflush size : 64
cache_alignment: 64
address sizes : 36 bits physical, 48 bits virtual
power management:processor : 1
vendor_id: GenuineIntel
cpu family : 6
model : 23
model name : Intel(R) Core(TM)2 Duo CPU E7500 @ 2.93GHz
stepping : 10
microcode :0xa0b
cpu MHz : 1603.000
cache size : 3072 KB
physical id : 0
siblings : 2
core id : 1
cpu cores : 2
apicid : 1
initial apicid : 1
fpu : yesfpu_exception: yes
cpuid level : 13
wp : yes
flags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca
cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx lmconstant_tscarch_perfmonpebs btsrep_goodnopl aperfmperf pni dtes64 monitords_cplvmx est tm2 ssse3 cx16 xtpr pdcmsse4_1xsavelahf_lmdthermtpr_shadowvnmi flexprioritybugs : bogomips : 5851.89 clflush size : 64
cache_alignment: 64
address sizes : 36 bits physical, 48 bits virtual
power management:[root@btrfs test_nosync_32768__sync_1_seqwr_4G_btrfs_1]#
Please tell me if you are interesting on more information or operation.
Thanks
ZhaoleiQuoted 7 lines
I remember an issue a few years ago where simply reverting a path that uninlined the
rw_semcode fixed a buffered I/O performance regression when using Samba on a very low end arm device, so everything is possible.I'd still like to ensure the numbers are reproducible in this case first, and look at all the information Jan asked for. Ask a next step we could then look at using an inline version to check if thast helps.