WSL: Clock sources are slow to read and imprecise
On a normal bare metal Linux system:
Name Cost(ns) +/- Resol Mono Fail Warp Stal Regr
tsc 10.97 0.21% ---- Yes 0 0 0 0
gettimeofday 17.60 0.01% 1000KHz No 0 0 1000 0
17.60 0.01%
realtime 17.60 0.07% ---- Yes 0 0 0 0
realtime_crs 4.66 0.06% 100Hz No 999 0 0 0
3.33 142.09%
monotonic 17.20 0.06% ---- Yes 0 0 0 0
monotonic_crs 4.52 0.02% 100Hz No 999 0 0 0
3.34 142.09%
monotonic_raw 69.76 0.10% ---- Yes 0 0 0 0
boottime 73.53 0.05% ---- Yes 0 0 0 0
process 115.67 0.08% ---- Yes 0 0 0 0
thread 109.64 0.07% ---- Yes 0 0 0 0
clock 116.93 0.04% 1000KHz No 0 0 1000 0
getrusage 181.32 0.05% 100Hz No 996 3 3 0
ftime 20.78 0.04% 1000Hz No 995 0 4 0
time 2.84 0.03% 1Hz No 1000 0 0 0
On “Bash on Windows”, the clock behavior isn’t as nice:
Name Cost(ns) +/- Resol Mono Fail Warp Stal Regr
tsc 12.53 86.77% ---- Yes 0 0 0 0
gettimeofday 166.10 0.72% 1032Hz No 943 5 55 0
realtime 182.39 1.02% 3333KHz No 0 0 131 0
realtime_crs 168.97 0.83% 2144Hz No 934 0 65 0
monotonic 186.90 0.38% 1000KHz No 0 0 998 0
monotonic_crs 188.76 0.33% 1000KHz No 0 0 998 0
Failed to read from monotonic_raw clock (1, 4)
boottime 186.66 0.37% 1000KHz No 0 0 998 0
process 272.19 0.81% 64Hz No 996 3 3 0
thread 229.25 0.84% 64Hz No 997 2 2 0
clock 278.50 1.23% 64Hz No 996 3 3 0
getrusage 385.45 0.78% 64Hz No 995 4 4 0
ftime 207.14 0.81% 1000Hz No 959 0 40 0
time 156.06 0.86% 1Hz No 1000 0 0 0
To explain the columns a bit:
The Cost(ns) column is the time cost to read that clock in nanoseconds, as measured by the TSC (except when measuring the cost of reading the TSC itself, which is measured by looking at gettimeofday).
Resol is the observable tick rate of the clock.
Mono indicates whether the clock source is monotonic, with an additional restriction. Not only must the clock only move forward, it must never return the same value (i.e. high frequency).
Fail indicates the number of times the clock source failed to advance in >= 200 reads.
Warp indicates the number of times the clock jumped forward by an unexpectedly large span (threshold for this is currently 1ms).
Stal indicates the number of times the clock source failed to advance in more than 1, but less than 200 reads.
Regr indicates the number of times the clock moved backwards. This can happen due to hypervisor behavior, NTP adjustments, and other clock changes. It’s extremely bad behavior if you use the clock source for timespan measurement (e.g. profiling, benchmarks, etc).
About this issue
- Original URL
- State: closed
- Created 11 years ago
- Reactions: 4
- Comments: 18 (2 by maintainers)
@russalex Understood. Just wanted to post an update with the latest build. Not trying to force a change ahead of schedule. 😄