Logger Benchmarks

Results

Integer Benchmark, One Thread

Library	50	75	90	95	99	99.9	Max	Version
xtr	11	12	12.5	13	14	41	55.5	2.0.1
platformlab_nanolog	11.5	11.5	12	12	12.5	14.5	33	85b71b6
quill	12	13	13.5	14	14.5	16	530.5	v2.0.2
reckless	17	17	17.5	17.5	19.5	21.5	38	v3.0.3
ms_binlog	29.5	31	32	32.5	68	105.5	384	2020-04-26-82-g2de2fa0
iyengar_nanolog	142	158	173.5	192.5	421	484	41490	3ffc74a
spdlog	284.5	317	346	363.5	399.5	501.5	572.5	v1.10.0
g3log	2543	2637	2724	2779	2879	3015	3354	1.3.4

Integer Benchmark, Four Threads

Library	50	75	90	95	99	99.9	Max	Version
xtr	11	12	12.5	13.5	14.5	41.5	73	2.0.1
platformlab_nanolog	11.5	12	12.5	12.5	13	17	33.5	85b71b6
quill	13	14.5	15.5	16.5	18	21.5	23	v2.0.2
reckless	18	19.5	20.5	20.5	21.5	28	65	v3.0.3
ms_binlog	29.5	31	32	33	70.5	107.5	412	2020-04-26-82-g2de2fa0
iyengar_nanolog	135	152.5	172	184	227	507.5	63290	3ffc74a
spdlog	280.5	314.5	342.5	361.5	403.5	540.5	669.5	v1.10.0
g3log	2530	2624	2710	2767	2878	3017	7146	1.3.4

String Benchmark, One Thread

Library	50	75	90	95	99	99.9	Max	Version
xtr	12.5	13.5	14	14.5	16	41.5	51	2.0.1
platformlab_nanolog	13.5	14	14.5	14.5	15.5	19	35.5	85b71b6
quill	14.5	15.5	18	18.5	20	21.5	22.5	v2.0.2
ms_binlog	34	35.5	36.5	38	75	109.5	389.5	2020-04-26-82-g2de2fa0
reckless	44	47.5	49.5	50.5	52.5	79.5	134.5	v3.0.3
iyengar_nanolog	120	150	166.5	181	379.5	492.5	41680	3ffc74a
spdlog	255	299	344	373.5	484.5	1123	1602	v1.10.0
g3log	1975	2214	2340	2416	2536	2693	2888	1.3.4

String Benchmark, Four Threads

Library	50	75	90	95	99	99.9	Max	Version
xtr	13	13.5	14.5	15	16.5	46	66	2.0.1
platformlab_nanolog	13.5	14	14.5	15	16.5	22	39	85b71b6
quill	14.5	15.5	17	18	20	21.5	23.5	v2.0.2
ms_binlog	34.5	36	37.5	38.5	75.5	115.5	408.5	2020-04-26-82-g2de2fa0
reckless	44	48	50	51	52.5	68.5	216.5	v3.0.3
iyengar_nanolog	96.5	132.5	161.5	174.5	229.5	487	8354	3ffc74a
spdlog	256	301.5	349.5	379	495.5	1177	2071	v1.10.0
g3log	1911	2245	2380	2451	2582	2738	6578	1.3.4

Note that these timings are not really percentiles, they are instead timings collected into buckets, the average of the bucket is recorded, then percentiles of the bucket averages are calculated.

Test Environment

• Stock AMD 5950X
• Samsung 970 EVO Plus 1TB SSD
• GCC 11.2.0
• Linux 5.17.0-1-amd64, Debian "bookworm"
• SMT disabled
• CPU governor set to 'performance'
• Kernel command line parameters:
  • mitigations=off
  • transparent_hugepage=never
  • isolcpus=1,2,3,4,5
  • nohz_full=1,2,3,4,5
  • rcu_nocbs=1,2,3,4,5
  • rcu_nocb_poll
  • processor.max_cstate=1
  • idle=poll
  • skew_tick=1
• Run irqbalance (IRQBALANCE_BANNED_CPULIST is not set as irqbalance will automatically add isolated and nohz cpus to the banned list).
• Run echo ffffffe1 > /sys/bus/workqueue/devices/writeback/cpumask to prevent bdi-flush threads from running on isolated cores.
• RCU thread affinities will be 0,6-15 on boot.
• amd_pstate scaling driver (this made no difference to the results, it is only listed here for completeness).

Changes From Upstream

• Reduced iterations from 100,000 to 10,000 as 10,000 was enough to get consistent results.
• The Reckless buffer size was increased to 512KB (default is 64KB).
• Affinity settings were applied to the Reckless background thread.
• The Iyengar roll size was increased to 10GB.
• PlatformLab's Config.h was modified; O_DSYNC was removed from from FILE_PARAMS, POLL_INTERVAL_NO_WORK_US was set to 0 and POLL_INTERVAL_DURING_IO_US was set to 0 (see platformlab.patch).
• PlatformLab's background thread has affinity set. This made a big improvement in performance---without this it would often run at around 30ns.
• Use cores 1-5 instead of 0-5 as core 0 is by default used as a 'housekeeping' core which cannot have nohz_full enabled.
• Turned on link time optimization.
• Other minor changes:
  • Changed static variables holding RNG state in the wait() function to thread_local as they are mutated by multiple threads.
  • TSC frequency is estimated once and cached.
  • Added fractions to timings.

Building

mkdir build
cd build
conan install ../third_party/xtr -g cmake_find_package
cmake .. -DCMAKE_MODULE_PATH=${PWD}