| Message ID | 20210409160305.711318-1-haris.iqbal@ionos.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | block: add two statistic tables | expand |
On 4/9/21 10:03 AM, Md Haris Iqbal wrote: > Hi Jens, > > This version fixes the long lines in the code as per Christoph's comment. I'd really like to see some solid justification for the addition, though. I clicked the v1 link and it's got details on what you get out of it, but not really the 'why' of reasoning for the feature. I mean, you could feasibly have a blktrace based userspace solution. Just wondering if that has been tried, I know that's what we do at Facebook for example.
On Fri, Apr 9, 2021 at 11:03 PM Jens Axboe <axboe@kernel.dk> wrote: > > On 4/9/21 10:03 AM, Md Haris Iqbal wrote: > > Hi Jens, > > > > This version fixes the long lines in the code as per Christoph's comment. > > I'd really like to see some solid justification for the addition, > though. I clicked the v1 link and it's got details on what you get out > of it, but not really the 'why' of reasoning for the feature. I mean, > you could feasibly have a blktrace based userspace solution. Just > wondering if that has been tried, I know that's what we do at Facebook > for example. > Hi Jens, Thanks for the reply. For the use case of the additional stats, as a cloud provider, we often need to handle report from the customers regarding performance problem in a period of time in the past, so it's not feasible for us to run blktrace, customer workload could change from time to time, with the additional stats, we gather through all metrics using Prometheus, we can navigate to the period of time interested, to check if the performance matches the SLA, it also helps us to find the user IO pattern, we can more easily reproduce. We do use blktrace from time to time too if it's not too late (when IO pattern has not changed.) Thanks! Jack
On Mon, Apr 12, 2021 at 7:35 AM Jinpu Wang <jinpu.wang@ionos.com> wrote: > > On Fri, Apr 9, 2021 at 11:03 PM Jens Axboe <axboe@kernel.dk> wrote: > > > > On 4/9/21 10:03 AM, Md Haris Iqbal wrote: > > > Hi Jens, > > > > > > This version fixes the long lines in the code as per Christoph's comment. > > > > I'd really like to see some solid justification for the addition, > > though. I clicked the v1 link and it's got details on what you get out > > of it, but not really the 'why' of reasoning for the feature. I mean, > > you could feasibly have a blktrace based userspace solution. Just > > wondering if that has been tried, I know that's what we do at Facebook > > for example. > > > Hi Jens, > > Thanks for the reply. > For the use case of the additional stats, as a cloud provider, we > often need to handle report from the customers regarding > performance problem in a period of time in the past, so it's not > feasible for us to run blktrace, customer workload could change from > time to time, with the additional stats, we gather through all metrics > using Prometheus, we can navigate to the period of time interested, > to check if the performance matches the SLA, it also helps us to find > the user IO pattern, we can more easily reproduce. > > We do use blktrace from time to time too if it's not too late (when IO > pattern has not changed.) > > Thanks! > Jack Hi Jens, A gentle ping!
On 4/11/21 11:35 PM, Jinpu Wang wrote: > On Fri, Apr 9, 2021 at 11:03 PM Jens Axboe <axboe@kernel.dk> wrote: >> >> On 4/9/21 10:03 AM, Md Haris Iqbal wrote: >>> Hi Jens, >>> >>> This version fixes the long lines in the code as per Christoph's comment. >> >> I'd really like to see some solid justification for the addition, >> though. I clicked the v1 link and it's got details on what you get out >> of it, but not really the 'why' of reasoning for the feature. I mean, >> you could feasibly have a blktrace based userspace solution. Just >> wondering if that has been tried, I know that's what we do at Facebook >> for example. >> > Hi Jens, > > Thanks for the reply. > For the use case of the additional stats, as a cloud provider, we > often need to handle report from the customers regarding > performance problem in a period of time in the past, so it's not > feasible for us to run blktrace, customer workload could change from > time to time, with the additional stats, we gather through all metrics > using Prometheus, we can navigate to the period of time interested, > to check if the performance matches the SLA, it also helps us to find > the user IO pattern, we can more easily reproduce. My suggestion isn't to run just blktrace all the time, rather collect the tracing info from there and store them away. Then you can go back in time and see what is going on. Hence my questioning on adding this new stat tracking, when it's already readily available to be consumed by a small daemon that can continually track it in userspace.
On Mon, Apr 19, 2021 at 7:57 PM Jens Axboe <axboe@kernel.dk> wrote: > > On 4/11/21 11:35 PM, Jinpu Wang wrote: > > On Fri, Apr 9, 2021 at 11:03 PM Jens Axboe <axboe@kernel.dk> wrote: > >> > >> On 4/9/21 10:03 AM, Md Haris Iqbal wrote: > >>> Hi Jens, > >>> > >>> This version fixes the long lines in the code as per Christoph's comment. > >> > >> I'd really like to see some solid justification for the addition, > >> though. I clicked the v1 link and it's got details on what you get out > >> of it, but not really the 'why' of reasoning for the feature. I mean, > >> you could feasibly have a blktrace based userspace solution. Just > >> wondering if that has been tried, I know that's what we do at Facebook > >> for example. > >> > > Hi Jens, > > > > Thanks for the reply. > > For the use case of the additional stats, as a cloud provider, we > > often need to handle report from the customers regarding > > performance problem in a period of time in the past, so it's not > > feasible for us to run blktrace, customer workload could change from > > time to time, with the additional stats, we gather through all metrics > > using Prometheus, we can navigate to the period of time interested, > > to check if the performance matches the SLA, it also helps us to find > > the user IO pattern, we can more easily reproduce. > > My suggestion isn't to run just blktrace all the time, rather collect > the tracing info from there and store them away. Then you can go back > in time and see what is going on. Hence my questioning on adding this > new stat tracking, when it's already readily available to be consumed > by a small daemon that can continually track it in userspace. > > -- > Jens Axboe > Hi Jens, The problem with using blktrace at production may cause a performance drop ~30%. while with the block stats here, we only see ~3% when enabled. We did a benchmark with rnbd. Fio config: [global] description=Emulation of Storage Server Access Pattern bssplit=512/20:1k/16:2k/9:4k/12:8k/19:16k/10:32k/8:64k/4 fadvise_hint=0 rw=randrw:2 direct=1 random_distribution=zipf:1.2 #size=1G time_based=1 runtime=10 ramp_time=1 ioengine=libaio iodepth=128 iodepth_batch_submit=128 iodepth_batch_complete=128 numjobs=1 #gtod_reduce=1 group_reporting [job1] filename=/dev/rnbd0 blktrace command: # blktrace -a read -a write -d /dev/rnbd0 read IOPS drops 35%, similar for write IOPS. RNBD-No-blktrace RNBD-With-blktrace 102056.894311 -35.5% The tests are done with v5.4.30. Test hardware is root@x4-left:~/haris/sds-perf# uname -a Linux x4-left 5.10.30-pserver #5.10.30-1+feature+linux+5.10.y+20210414.1233+e3dd267~deb10 SMP x86_64 GNU/Linux root@x4-left:~/haris/sds-perf# lscpu Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian Address sizes: 46 bits physical, 48 bits virtual CPU(s): 40 On-line CPU(s) list: 0-39 Thread(s) per core: 2 Core(s) per socket: 10 Socket(s): 2 NUMA node(s): 2 Vendor ID: GenuineIntel CPU family: 6 Model: 85 Model name: Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz Stepping: 4 CPU MHz: 800.571 CPU max MHz: 3000.0000 CPU min MHz: 800.0000 BogoMIPS: 4400.00 Virtualization: VT-x L1d cache: 32K L1i cache: 32K L2 cache: 1024K L3 cache: 14080K NUMA node0 CPU(s): 0-9,20-29 NUMA node1 CPU(s): 10-19,30-39 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 pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cdp_l3 invpcid_single intel_ppin mba ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts pku ospke root@x4-left:~/haris/sds-perf# ibstat CA 'mlx5_0' CA type: MT4115 Number of ports: 1 Firmware version: 12.26.4012 Hardware version: 0 Node GUID: 0xec0d9a0300c5fffc System image GUID: 0xec0d9a0300c5fffc Port 1: State: Active Physical state: LinkUp Rate: 100 Base lid: 3 LMC: 0 SM lid: 3 Capability mask: 0x2651e84a Port GUID: 0xec0d9a0300c5fffc Link layer: InfiniBand CA 'mlx5_1' CA type: MT4115 Number of ports: 1 Firmware version: 12.26.4012 Hardware version: 0 Node GUID: 0xec0d9a0300c5fffd System image GUID: 0xec0d9a0300c5fffc Port 1: State: Active Physical state: LinkUp Rate: 100 Base lid: 1 LMC: 0 SM lid: 1 Capability mask: 0x2651e84a Port GUID: 0xec0d9a0300c5fffd Link layer: InfiniBand Thanks!
On 4/21/21 8:49 AM, Jinpu Wang wrote: > On Mon, Apr 19, 2021 at 7:57 PM Jens Axboe <axboe@kernel.dk> wrote: >> >> On 4/11/21 11:35 PM, Jinpu Wang wrote: >>> On Fri, Apr 9, 2021 at 11:03 PM Jens Axboe <axboe@kernel.dk> wrote: >>>> >>>> On 4/9/21 10:03 AM, Md Haris Iqbal wrote: >>>>> Hi Jens, >>>>> >>>>> This version fixes the long lines in the code as per Christoph's comment. >>>> >>>> I'd really like to see some solid justification for the addition, >>>> though. I clicked the v1 link and it's got details on what you get out >>>> of it, but not really the 'why' of reasoning for the feature. I mean, >>>> you could feasibly have a blktrace based userspace solution. Just >>>> wondering if that has been tried, I know that's what we do at Facebook >>>> for example. >>>> >>> Hi Jens, >>> >>> Thanks for the reply. >>> For the use case of the additional stats, as a cloud provider, we >>> often need to handle report from the customers regarding >>> performance problem in a period of time in the past, so it's not >>> feasible for us to run blktrace, customer workload could change from >>> time to time, with the additional stats, we gather through all metrics >>> using Prometheus, we can navigate to the period of time interested, >>> to check if the performance matches the SLA, it also helps us to find >>> the user IO pattern, we can more easily reproduce. >> >> My suggestion isn't to run just blktrace all the time, rather collect >> the tracing info from there and store them away. Then you can go back >> in time and see what is going on. Hence my questioning on adding this >> new stat tracking, when it's already readily available to be consumed >> by a small daemon that can continually track it in userspace. >> >> -- >> Jens Axboe >> > Hi Jens, > The problem with using blktrace at production may cause a performance > drop ~30%. while with the block stats here, we only see ~3% when > enabled. It's probably was asked before, but let's refresh as the discussion erupted again. I get your problem with blktrace(8), IIRC it definitely can deteriorate performance if run constantly, but did you try to write a bpf program that does smarter accumulation in the kernel? Like making bpf to collect a latency table (right as in your patches do) and flushing it to the disk periodically? > We did a benchmark with rnbd. > > Fio config: > [global] > description=Emulation of Storage Server Access Pattern > bssplit=512/20:1k/16:2k/9:4k/12:8k/19:16k/10:32k/8:64k/4 > fadvise_hint=0 > rw=randrw:2 > direct=1 > random_distribution=zipf:1.2 > #size=1G > time_based=1 > runtime=10 > ramp_time=1 > ioengine=libaio > iodepth=128 > iodepth_batch_submit=128 > iodepth_batch_complete=128 > numjobs=1 > #gtod_reduce=1 > group_reporting > > [job1] > filename=/dev/rnbd0 > > blktrace command: > # blktrace -a read -a write -d /dev/rnbd0 > > read IOPS drops 35%, similar for write IOPS. > > RNBD-No-blktrace RNBD-With-blktrace > > 102056.894311 -35.5% > > The tests are done with v5.4.30. > Test hardware is > root@x4-left:~/haris/sds-perf# uname -a > Linux x4-left 5.10.30-pserver > #5.10.30-1+feature+linux+5.10.y+20210414.1233+e3dd267~deb10 SMP > x86_64 GNU/Linux > root@x4-left:~/haris/sds-perf# lscpu > Architecture: x86_64 > CPU op-mode(s): 32-bit, 64-bit > Byte Order: Little Endian > Address sizes: 46 bits physical, 48 bits virtual > CPU(s): 40 > On-line CPU(s) list: 0-39 > Thread(s) per core: 2 > Core(s) per socket: 10 > Socket(s): 2 > NUMA node(s): 2 > Vendor ID: GenuineIntel > CPU family: 6 > Model: 85 > Model name: Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz > Stepping: 4 > CPU MHz: 800.571 > CPU max MHz: 3000.0000 > CPU min MHz: 800.0000 > BogoMIPS: 4400.00 > Virtualization: VT-x > L1d cache: 32K > L1i cache: 32K > L2 cache: 1024K > L3 cache: 14080K > NUMA node0 CPU(s): 0-9,20-29 > NUMA node1 CPU(s): 10-19,30-39 > 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 pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts > rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq > dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm > pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes > xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 > cdp_l3 invpcid_single intel_ppin mba ibrs ibpb stibp tpr_shadow vnmi > flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep > bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap > clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec > xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local > dtherm ida arat pln pts pku ospke > root@x4-left:~/haris/sds-perf# ibstat > CA 'mlx5_0' > CA type: MT4115 > Number of ports: 1 > Firmware version: 12.26.4012 > Hardware version: 0 > Node GUID: 0xec0d9a0300c5fffc > System image GUID: 0xec0d9a0300c5fffc > Port 1: > State: Active > Physical state: LinkUp > Rate: 100 > Base lid: 3 > LMC: 0 > SM lid: 3 > Capability mask: 0x2651e84a > Port GUID: 0xec0d9a0300c5fffc > Link layer: InfiniBand > CA 'mlx5_1' > CA type: MT4115 > Number of ports: 1 > Firmware version: 12.26.4012 > Hardware version: 0 > Node GUID: 0xec0d9a0300c5fffd > System image GUID: 0xec0d9a0300c5fffc > Port 1: > State: Active > Physical state: LinkUp > Rate: 100 > Base lid: 1 > LMC: 0 > SM lid: 1 > Capability mask: 0x2651e84a > Port GUID: 0xec0d9a0300c5fffd > Link layer: InfiniBand > > Thanks! >
On Wed, Apr 21, 2021 at 12:20 PM Pavel Begunkov <asml.silence@gmail.com> wrote: > > On 4/21/21 8:49 AM, Jinpu Wang wrote: > > On Mon, Apr 19, 2021 at 7:57 PM Jens Axboe <axboe@kernel.dk> wrote: > >> > >> On 4/11/21 11:35 PM, Jinpu Wang wrote: > >>> On Fri, Apr 9, 2021 at 11:03 PM Jens Axboe <axboe@kernel.dk> wrote: > >>>> > >>>> On 4/9/21 10:03 AM, Md Haris Iqbal wrote: > >>>>> Hi Jens, > >>>>> > >>>>> This version fixes the long lines in the code as per Christoph's comment. > >>>> > >>>> I'd really like to see some solid justification for the addition, > >>>> though. I clicked the v1 link and it's got details on what you get out > >>>> of it, but not really the 'why' of reasoning for the feature. I mean, > >>>> you could feasibly have a blktrace based userspace solution. Just > >>>> wondering if that has been tried, I know that's what we do at Facebook > >>>> for example. > >>>> > >>> Hi Jens, > >>> > >>> Thanks for the reply. > >>> For the use case of the additional stats, as a cloud provider, we > >>> often need to handle report from the customers regarding > >>> performance problem in a period of time in the past, so it's not > >>> feasible for us to run blktrace, customer workload could change from > >>> time to time, with the additional stats, we gather through all metrics > >>> using Prometheus, we can navigate to the period of time interested, > >>> to check if the performance matches the SLA, it also helps us to find > >>> the user IO pattern, we can more easily reproduce. > >> > >> My suggestion isn't to run just blktrace all the time, rather collect > >> the tracing info from there and store them away. Then you can go back > >> in time and see what is going on. Hence my questioning on adding this > >> new stat tracking, when it's already readily available to be consumed > >> by a small daemon that can continually track it in userspace. > >> > >> -- > >> Jens Axboe > >> > > Hi Jens, > > The problem with using blktrace at production may cause a performance > > drop ~30%. while with the block stats here, we only see ~3% when > > enabled. > > It's probably was asked before, but let's refresh as the discussion > erupted again. > > I get your problem with blktrace(8), IIRC it definitely can deteriorate > performance if run constantly, but did you try to write a bpf program > that does smarter accumulation in the kernel? Like making bpf to collect > a latency table (right as in your patches do) and flushing it to the > disk periodically? Hi Pavel, Thanks for the suggestion. We did test with ebpf with kprobe in the past (~kernel 4.4/4.14), we saw 10% performance drop, that's the reason we develop this stats patches. But I just did another test with bpftrace on k 5.10.30, I do not see performance lost. It must be ebpf is improving very much since then. So to summarize, we can use bpftrace to do the drop in latest kernel, there is no need to have it build into the kernel. Thanks! the bpftrace I used during testing: root@x4-left:~# cat /usr/sbin/biolatency.bt #!/usr/bin/env bpftrace /* * biolatency.bt Block I/O latency as a histogram. * For Linux, uses bpftrace, eBPF. * * This is a bpftrace version of the bcc tool of the same name. * * Copyright 2018 Netflix, Inc. * Licensed under the Apache License, Version 2.0 (the "License") * * 13-Sep-2018 Brendan Gregg Created this. */ BEGIN { printf("Tracing block device I/O... Hit Ctrl-C to end.\n"); } kprobe:blk_account_io_start { @start[arg0] = nsecs; } kprobe:blk_account_io_done /@start[arg0]/ { @usecs = hist((nsecs - @start[arg0]) / 1000); delete(@start[arg0]); } > > > > We did a benchmark with rnbd. > > > > Fio config: > > [global] > > description=Emulation of Storage Server Access Pattern > > bssplit=512/20:1k/16:2k/9:4k/12:8k/19:16k/10:32k/8:64k/4 > > fadvise_hint=0 > > rw=randrw:2 > > direct=1 > > random_distribution=zipf:1.2 > > #size=1G > > time_based=1 > > runtime=10 > > ramp_time=1 > > ioengine=libaio > > iodepth=128 > > iodepth_batch_submit=128 > > iodepth_batch_complete=128 > > numjobs=1 > > #gtod_reduce=1 > > group_reporting > > > > [job1] > > filename=/dev/rnbd0 > > > > blktrace command: > > # blktrace -a read -a write -d /dev/rnbd0 > > > > read IOPS drops 35%, similar for write IOPS. > > > > RNBD-No-blktrace RNBD-With-blktrace > > > > 102056.894311 -35.5% > > > > The tests are done with v5.4.30. > > Test hardware is > > root@x4-left:~/haris/sds-perf# uname -a > > Linux x4-left 5.10.30-pserver > > #5.10.30-1+feature+linux+5.10.y+20210414.1233+e3dd267~deb10 SMP > > x86_64 GNU/Linux > > root@x4-left:~/haris/sds-perf# lscpu > > Architecture: x86_64 > > CPU op-mode(s): 32-bit, 64-bit > > Byte Order: Little Endian > > Address sizes: 46 bits physical, 48 bits virtual > > CPU(s): 40 > > On-line CPU(s) list: 0-39 > > Thread(s) per core: 2 > > Core(s) per socket: 10 > > Socket(s): 2 > > NUMA node(s): 2 > > Vendor ID: GenuineIntel > > CPU family: 6 > > Model: 85 > > Model name: Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz > > Stepping: 4 > > CPU MHz: 800.571 > > CPU max MHz: 3000.0000 > > CPU min MHz: 800.0000 > > BogoMIPS: 4400.00 > > Virtualization: VT-x > > L1d cache: 32K > > L1i cache: 32K > > L2 cache: 1024K > > L3 cache: 14080K > > NUMA node0 CPU(s): 0-9,20-29 > > NUMA node1 CPU(s): 10-19,30-39 > > 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 pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts > > rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq > > dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm > > pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes > > xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 > > cdp_l3 invpcid_single intel_ppin mba ibrs ibpb stibp tpr_shadow vnmi > > flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep > > bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap > > clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec > > xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local > > dtherm ida arat pln pts pku ospke > > root@x4-left:~/haris/sds-perf# ibstat > > CA 'mlx5_0' > > CA type: MT4115 > > Number of ports: 1 > > Firmware version: 12.26.4012 > > Hardware version: 0 > > Node GUID: 0xec0d9a0300c5fffc > > System image GUID: 0xec0d9a0300c5fffc > > Port 1: > > State: Active > > Physical state: LinkUp > > Rate: 100 > > Base lid: 3 > > LMC: 0 > > SM lid: 3 > > Capability mask: 0x2651e84a > > Port GUID: 0xec0d9a0300c5fffc > > Link layer: InfiniBand > > CA 'mlx5_1' > > CA type: MT4115 > > Number of ports: 1 > > Firmware version: 12.26.4012 > > Hardware version: 0 > > Node GUID: 0xec0d9a0300c5fffd > > System image GUID: 0xec0d9a0300c5fffc > > Port 1: > > State: Active > > Physical state: LinkUp > > Rate: 100 > > Base lid: 1 > > LMC: 0 > > SM lid: 1 > > Capability mask: 0x2651e84a > > Port GUID: 0xec0d9a0300c5fffd > > Link layer: InfiniBand > > > > Thanks! > > > > -- > Pavel Begunkov
On 4/21/21 12:50 PM, Jinpu Wang wrote: [snip] >>> Hi Jens, >>> The problem with using blktrace at production may cause a performance >>> drop ~30%. while with the block stats here, we only see ~3% when >>> enabled. >> >> It's probably was asked before, but let's refresh as the discussion >> erupted again. >> >> I get your problem with blktrace(8), IIRC it definitely can deteriorate >> performance if run constantly, but did you try to write a bpf program >> that does smarter accumulation in the kernel? Like making bpf to collect >> a latency table (right as in your patches do) and flushing it to the >> disk periodically? > Hi Pavel, > > Thanks for the suggestion. > > We did test with ebpf with kprobe in the past (~kernel 4.4/4.14), we > saw 10% performance drop, that's the reason we develop this > stats patches. > > But I just did another test with bpftrace on k 5.10.30, I do not see > performance lost. > It must be ebpf is improving very much since then. > > So to summarize, we can use bpftrace to do the drop in latest kernel, > there is no need to have it build into the kernel. Perfect, and I'm sure it will be even more convenient for you, for instance to gather other stats or do it somehow differently > Thanks! > > the bpftrace I used during testing: > root@x4-left:~# cat /usr/sbin/biolatency.bt > #!/usr/bin/env bpftrace > /* > * biolatency.bt Block I/O latency as a histogram. > * For Linux, uses bpftrace, eBPF. > * > * This is a bpftrace version of the bcc tool of the same name. > * > * Copyright 2018 Netflix, Inc. > * Licensed under the Apache License, Version 2.0 (the "License") > * > * 13-Sep-2018 Brendan Gregg Created this. > */ > > BEGIN > { > printf("Tracing block device I/O... Hit Ctrl-C to end.\n"); > } > > kprobe:blk_account_io_start > { > @start[arg0] = nsecs; > } > > kprobe:blk_account_io_done > /@start[arg0]/ > > { > @usecs = hist((nsecs - @start[arg0]) / 1000); > delete(@start[arg0]); > } > > > >> >> >>> We did a benchmark with rnbd. >>> >>> Fio config: >>> [global] >>> description=Emulation of Storage Server Access Pattern >>> bssplit=512/20:1k/16:2k/9:4k/12:8k/19:16k/10:32k/8:64k/4 >>> fadvise_hint=0 >>> rw=randrw:2 >>> direct=1 >>> random_distribution=zipf:1.2 >>> #size=1G >>> time_based=1 >>> runtime=10 >>> ramp_time=1 >>> ioengine=libaio >>> iodepth=128 >>> iodepth_batch_submit=128 >>> iodepth_batch_complete=128 >>> numjobs=1 >>> #gtod_reduce=1 >>> group_reporting >>> >>> [job1] >>> filename=/dev/rnbd0 >>> >>> blktrace command: >>> # blktrace -a read -a write -d /dev/rnbd0 >>> >>> read IOPS drops 35%, similar for write IOPS. >>> >>> RNBD-No-blktrace RNBD-With-blktrace >>> >>> 102056.894311 -35.5% >>> >>> The tests are done with v5.4.30. >>> Test hardware is >>> root@x4-left:~/haris/sds-perf# uname -a >>> Linux x4-left 5.10.30-pserver >>> #5.10.30-1+feature+linux+5.10.y+20210414.1233+e3dd267~deb10 SMP >>> x86_64 GNU/Linux >>> root@x4-left:~/haris/sds-perf# lscpu >>> Architecture: x86_64 >>> CPU op-mode(s): 32-bit, 64-bit >>> Byte Order: Little Endian >>> Address sizes: 46 bits physical, 48 bits virtual >>> CPU(s): 40 >>> On-line CPU(s) list: 0-39 >>> Thread(s) per core: 2 >>> Core(s) per socket: 10 >>> Socket(s): 2 >>> NUMA node(s): 2 >>> Vendor ID: GenuineIntel >>> CPU family: 6 >>> Model: 85 >>> Model name: Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz >>> Stepping: 4 >>> CPU MHz: 800.571 >>> CPU max MHz: 3000.0000 >>> CPU min MHz: 800.0000 >>> BogoMIPS: 4400.00 >>> Virtualization: VT-x >>> L1d cache: 32K >>> L1i cache: 32K >>> L2 cache: 1024K >>> L3 cache: 14080K >>> NUMA node0 CPU(s): 0-9,20-29 >>> NUMA node1 CPU(s): 10-19,30-39 >>> 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 pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts >>> rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq >>> dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm >>> pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes >>> xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 >>> cdp_l3 invpcid_single intel_ppin mba ibrs ibpb stibp tpr_shadow vnmi >>> flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep >>> bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap >>> clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec >>> xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local >>> dtherm ida arat pln pts pku ospke >>> root@x4-left:~/haris/sds-perf# ibstat >>> CA 'mlx5_0' >>> CA type: MT4115 >>> Number of ports: 1 >>> Firmware version: 12.26.4012 >>> Hardware version: 0 >>> Node GUID: 0xec0d9a0300c5fffc >>> System image GUID: 0xec0d9a0300c5fffc >>> Port 1: >>> State: Active >>> Physical state: LinkUp >>> Rate: 100 >>> Base lid: 3 >>> LMC: 0 >>> SM lid: 3 >>> Capability mask: 0x2651e84a >>> Port GUID: 0xec0d9a0300c5fffc >>> Link layer: InfiniBand >>> CA 'mlx5_1' >>> CA type: MT4115 >>> Number of ports: 1 >>> Firmware version: 12.26.4012 >>> Hardware version: 0 >>> Node GUID: 0xec0d9a0300c5fffd >>> System image GUID: 0xec0d9a0300c5fffc >>> Port 1: >>> State: Active >>> Physical state: LinkUp >>> Rate: 100 >>> Base lid: 1 >>> LMC: 0 >>> SM lid: 1 >>> Capability mask: 0x2651e84a >>> Port GUID: 0xec0d9a0300c5fffd >>> Link layer: InfiniBand >>> >>> Thanks!
On Wed, Apr 21, 2021 at 1:55 PM Pavel Begunkov <asml.silence@gmail.com> wrote: > > On 4/21/21 12:50 PM, Jinpu Wang wrote: > [snip] > >>> Hi Jens, > >>> The problem with using blktrace at production may cause a performance > >>> drop ~30%. while with the block stats here, we only see ~3% when > >>> enabled. > >> > >> It's probably was asked before, but let's refresh as the discussion > >> erupted again. > >> > >> I get your problem with blktrace(8), IIRC it definitely can deteriorate > >> performance if run constantly, but did you try to write a bpf program > >> that does smarter accumulation in the kernel? Like making bpf to collect > >> a latency table (right as in your patches do) and flushing it to the > >> disk periodically? > > Hi Pavel, > > > > Thanks for the suggestion. > > > > We did test with ebpf with kprobe in the past (~kernel 4.4/4.14), we > > saw 10% performance drop, that's the reason we develop this > > stats patches. > > > > But I just did another test with bpftrace on k 5.10.30, I do not see > > performance lost. > > It must be ebpf is improving very much since then. > > > > So to summarize, we can use bpftrace to do the drop in latest kernel, > > there is no need to have it build into the kernel. > > Perfect, and I'm sure it will be even more convenient for you, for > instance to gather other stats or do it somehow differently Yeah, agree. Thanks again! > > > Thanks! > > > > the bpftrace I used during testing: > > root@x4-left:~# cat /usr/sbin/biolatency.bt > > #!/usr/bin/env bpftrace > > /* > > * biolatency.bt Block I/O latency as a histogram. > > * For Linux, uses bpftrace, eBPF. > > * > > * This is a bpftrace version of the bcc tool of the same name. > > * > > * Copyright 2018 Netflix, Inc. > > * Licensed under the Apache License, Version 2.0 (the "License") > > * > > * 13-Sep-2018 Brendan Gregg Created this. > > */ > > > > BEGIN > > { > > printf("Tracing block device I/O... Hit Ctrl-C to end.\n"); > > } > > > > kprobe:blk_account_io_start > > { > > @start[arg0] = nsecs; > > } > > > > kprobe:blk_account_io_done > > /@start[arg0]/ > > > > { > > @usecs = hist((nsecs - @start[arg0]) / 1000); > > delete(@start[arg0]); > > } > > > > > > > >> > >> > >>> We did a benchmark with rnbd. > >>> > >>> Fio config: > >>> [global] > >>> description=Emulation of Storage Server Access Pattern > >>> bssplit=512/20:1k/16:2k/9:4k/12:8k/19:16k/10:32k/8:64k/4 > >>> fadvise_hint=0 > >>> rw=randrw:2 > >>> direct=1 > >>> random_distribution=zipf:1.2 > >>> #size=1G > >>> time_based=1 > >>> runtime=10 > >>> ramp_time=1 > >>> ioengine=libaio > >>> iodepth=128 > >>> iodepth_batch_submit=128 > >>> iodepth_batch_complete=128 > >>> numjobs=1 > >>> #gtod_reduce=1 > >>> group_reporting > >>> > >>> [job1] > >>> filename=/dev/rnbd0 > >>> > >>> blktrace command: > >>> # blktrace -a read -a write -d /dev/rnbd0 > >>> > >>> read IOPS drops 35%, similar for write IOPS. > >>> > >>> RNBD-No-blktrace RNBD-With-blktrace > >>> > >>> 102056.894311 -35.5% > >>> > >>> The tests are done with v5.4.30. > >>> Test hardware is > >>> root@x4-left:~/haris/sds-perf# uname -a > >>> Linux x4-left 5.10.30-pserver > >>> #5.10.30-1+feature+linux+5.10.y+20210414.1233+e3dd267~deb10 SMP > >>> x86_64 GNU/Linux > >>> root@x4-left:~/haris/sds-perf# lscpu > >>> Architecture: x86_64 > >>> CPU op-mode(s): 32-bit, 64-bit > >>> Byte Order: Little Endian > >>> Address sizes: 46 bits physical, 48 bits virtual > >>> CPU(s): 40 > >>> On-line CPU(s) list: 0-39 > >>> Thread(s) per core: 2 > >>> Core(s) per socket: 10 > >>> Socket(s): 2 > >>> NUMA node(s): 2 > >>> Vendor ID: GenuineIntel > >>> CPU family: 6 > >>> Model: 85 > >>> Model name: Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz > >>> Stepping: 4 > >>> CPU MHz: 800.571 > >>> CPU max MHz: 3000.0000 > >>> CPU min MHz: 800.0000 > >>> BogoMIPS: 4400.00 > >>> Virtualization: VT-x > >>> L1d cache: 32K > >>> L1i cache: 32K > >>> L2 cache: 1024K > >>> L3 cache: 14080K > >>> NUMA node0 CPU(s): 0-9,20-29 > >>> NUMA node1 CPU(s): 10-19,30-39 > >>> 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 pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts > >>> rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq > >>> dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm > >>> pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes > >>> xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 > >>> cdp_l3 invpcid_single intel_ppin mba ibrs ibpb stibp tpr_shadow vnmi > >>> flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep > >>> bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap > >>> clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec > >>> xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local > >>> dtherm ida arat pln pts pku ospke > >>> root@x4-left:~/haris/sds-perf# ibstat > >>> CA 'mlx5_0' > >>> CA type: MT4115 > >>> Number of ports: 1 > >>> Firmware version: 12.26.4012 > >>> Hardware version: 0 > >>> Node GUID: 0xec0d9a0300c5fffc > >>> System image GUID: 0xec0d9a0300c5fffc > >>> Port 1: > >>> State: Active > >>> Physical state: LinkUp > >>> Rate: 100 > >>> Base lid: 3 > >>> LMC: 0 > >>> SM lid: 3 > >>> Capability mask: 0x2651e84a > >>> Port GUID: 0xec0d9a0300c5fffc > >>> Link layer: InfiniBand > >>> CA 'mlx5_1' > >>> CA type: MT4115 > >>> Number of ports: 1 > >>> Firmware version: 12.26.4012 > >>> Hardware version: 0 > >>> Node GUID: 0xec0d9a0300c5fffd > >>> System image GUID: 0xec0d9a0300c5fffc > >>> Port 1: > >>> State: Active > >>> Physical state: LinkUp > >>> Rate: 100 > >>> Base lid: 1 > >>> LMC: 0 > >>> SM lid: 1 > >>> Capability mask: 0x2651e84a > >>> Port GUID: 0xec0d9a0300c5fffd > >>> Link layer: InfiniBand > >>> > >>> Thanks! > > -- > Pavel Begunkov
On 4/21/21 5:57 AM, Jinpu Wang wrote: > On Wed, Apr 21, 2021 at 1:55 PM Pavel Begunkov <asml.silence@gmail.com> wrote: >> >> On 4/21/21 12:50 PM, Jinpu Wang wrote: >> [snip] >>>>> Hi Jens, >>>>> The problem with using blktrace at production may cause a performance >>>>> drop ~30%. while with the block stats here, we only see ~3% when >>>>> enabled. >>>> >>>> It's probably was asked before, but let's refresh as the discussion >>>> erupted again. >>>> >>>> I get your problem with blktrace(8), IIRC it definitely can deteriorate >>>> performance if run constantly, but did you try to write a bpf program >>>> that does smarter accumulation in the kernel? Like making bpf to collect >>>> a latency table (right as in your patches do) and flushing it to the >>>> disk periodically? >>> Hi Pavel, >>> >>> Thanks for the suggestion. >>> >>> We did test with ebpf with kprobe in the past (~kernel 4.4/4.14), we >>> saw 10% performance drop, that's the reason we develop this >>> stats patches. >>> >>> But I just did another test with bpftrace on k 5.10.30, I do not see >>> performance lost. >>> It must be ebpf is improving very much since then. >>> >>> So to summarize, we can use bpftrace to do the drop in latest kernel, >>> there is no need to have it build into the kernel. >> >> Perfect, and I'm sure it will be even more convenient for you, for >> instance to gather other stats or do it somehow differently > > Yeah, agree. > Thanks again! Perfect, thanks! I'll drop the series.
Hi Jens, This version fixes the long lines in the code as per Christoph's comment. Thanks, Haris PATCH V5: https://lore.kernel.org/linux-block/20210408135840.386076-1-haris.iqbal@ionos.com/ * Rebased with latest code. PATCH V4: https://lore.kernel.org/linux-block/20210203151019.27036-1-guoqing.jiang@cloud.ionos.com/ * Adds Reviewed-by tag from Johannes. PATCH V3: https://lore.kernel.org/linux-block/7f78132a-affc-eb03-735a-4da43e143b6e@cloud.ionos.com/T/#t * reorgnize the patchset per Johannes's suggestion. PATCH V2: https://lore.kernel.org/linux-block/20210201012727.28305-1-guoqing.jiang@cloud.ionos.com/T/#t *. remove BLK_ADDITIONAL_DISKSTAT option per Christoph's comment. *. move blk_queue_io_extra_stat into blk_additional_{latency,sector} per Christoph's comment. *. simplify blk_additional_latency by pass duration time directly. PATCH V1: https://marc.info/?l=linux-block&m=161176000024443&w=2 * add Jack's reviewed-by. RFC V5: https://marc.info/?l=linux-block&m=161789033303172&w=2 * Fix long lines per Christoph's comment. RFC V4: https://marc.info/?l=linux-block&m=161027198729158&w=2 * rebase with latest code. RFC V3: https://marc.info/?l=linux-block&m=159730633416534&w=2 * Move the #ifdef CONFIG_BLK_ADDITIONAL_DISKSTAT into the function body per Johannes's comment. * Tweak the output of two tables to make they are more intuitive RFC V2: https://marc.info/?l=linux-block&m=159467483514062&w=2 * don't call ktime_get_ns and drop unnecessary patches. * add io_extra_stats to avoid potential overhead. RFC V1: https://marc.info/?l=linux-block&m=159419516730386&w=2 Guoqing Jiang (3): block: add io_extra_stats node block: add a statistic table for io latency block: add a statistic table for io sector Documentation/ABI/testing/sysfs-block | 26 +++++++++ Documentation/block/queue-sysfs.rst | 5 ++ block/blk-core.c | 44 +++++++++++++++ block/blk-sysfs.c | 3 + block/genhd.c | 81 +++++++++++++++++++++++++++ include/linux/blkdev.h | 2 + include/linux/part_stat.h | 6 ++ 7 files changed, 167 insertions(+)