Golang library for access the CPU timestamp cycle counter (TSC) on x86-64. If not familar with using the TSC for benchmarking, refer to the Intel whitepaper. This is designed to be used for benchmarking code, so takes steps to prevent instruction reordering across measurement boundaries by the CPU.
Golang 1.4 or later is currently supported and x86-64 architetcture. The package will build on other architectures but all functions will simply return 0.
package main
import (
"fmt"
"github.com/dterei/gotsc"
)
const N = 100
func main() {
tsc := gotsc.TSCOverhead()
fmt.Println("TSC Overhead:", tsc)
start := gotsc.BenchStart()
for i := 0; i < N; i++ {
// code to evaluate
}
end := gotsc.BenchEnd()
avg := (end - start - tsc) / N
fmt.Println("Cycles:", avg)
}
There are two advantages over the standard golang time.Now()
function:
- Measurement is in cycles - for many situations cycle count is a more informative number than wall-clock time.
- Careful use of CPU serializing instructions to ensure no code you are benchmarking is moved outside the timed region, and no code you aren't benchmarking is moved into it.
Claim (2) may be a little contensious, so see below. For benchmarking with the TSC, we use the approach suggested by Intel:
cpuid
rdtsc
// code to benchmark
rdtscp
cpuid
There appears to be only confusion on what is both the correct and best way to read the TSC when benchmarking code. The most obvious and naive approach would be:
// code before
rdtsc
// code to benchmark
rdtsc
// code after
But rdtsc
doesn't prevent instructions being reordered by the CPU around it.
Thus, code before and after could move into the benchmarked region, while code
in the benchmarked region could move out.
The best Intel documentation on this suggests the following approach:
// code before
cpuid
rdtsc
// code to benchmark
rdtscp
cpuid
/// code after
The cpuid
instruction is a full barrier, preventing reordering in both
directions, while rdtscp
prevents reordering from above. We use rdtscp
at
the end rather than cpuid; rdtsc
as cpuid
is an expensive instruction with
high variance, so we want it outside the benchmarked region.
Ideally our benchmarking approach provides the following:
- Low variance for instructions involved in retrieving start and end TSC so that we can subtract their overhead from the measurement with more confidence.
- Low cost to read the TSC so that we can take benchmarks as often as possible without affecting application performance.
- High resolution so that we can measure the cost of very small sets of instructions.
The recommended Intel approach provides (1) and (3) but the use of cpuid
is
fairly expensive. The Intel SDM suggest that the lfence
instruction can be
used as an alternative to cpuid
, while AMD suggest the use of mfence
.
Linux takes this approach. Originally (LKML), it used an
lfence
either side with the thinking being that lfence
only prevents
reordering from above:
// Linux kernel TSC usage (circa 2008)
lfence
rdtsc
lfence
This was later (LKML) 'optimized' to just use one lfence
before rdtsc
:
// Linux kernel TSC usage (circa 2011+)
lfence
rdtsc
The kernel developer of the older lfence
both sides approach appears to
object to this optimization as 'unsafe' due to being a barrier in only one
direction. The 'modern' thinking appears to be that while this is technically
true, a microprocessor would never take advantage of this reordering---there is
no performance reason to do so.
The Akaros project investigated a number of alternative approaches (including all the above issues). Eventually taking the modern Linux approach and suggesting the following:
// code before
lfence
rdtsc
// code to benchmark
lfence
rdtsc
// code after
Finally, for complete reference, an older Intel guide to using
rdtsc
(pre rdtscp
days) suggest that you 'warm up' the cpuid
and rdtsc
instructions a few times before benchmarking the code:
// code before
// warmup
cpuid
rdtsc
cpuid
rdtsc
cpuid
rdtsc
cpuid
rdtsc
// code to benchmark
cpuid
rdtsc
// code afer
It's not clear if this is valuable any more when we have rdtscp
to avoid
include the highly variable cpuid
in our measurement region.
This is a very confusing situation. We keep it simple and stick with the
recommendation from Intel. This works well, but is a little more expensive due
to the cpuid
calls compared to alternatives. However, it also appears to be
the 'safest', ensuring accurate measurements. For very frequent calls to the
TSC when benchmarking is not your goal, the standard Go time.Now()
call is
very fast, essentially being lfence; rdtsc
.
To convert from cycles to wall-clock time we need to know TSC frequency. Frequency scaling on modern Intel chips doesn't affect the TSC.
Sadly, the only way to determine the TSC frequency appears to be through a MSR
using the rdmsr
instruction. This instruction is privileged and can't be
executed from user-space.
If we could, we want to access the MSR_PLATFORM_INFO
:
Register Name: MSR_PLATFORM_INFO [15:8] Description: Package Maximum Non-Turbo Ratio (R/O) The is the ratio of the frequency that invariant TSC runs at. Frequency = ratio * 100 MHz.
The multiplicative factor of 100 MHz
varies across architectures. Luckily, it
appears to be 100 MHz
on all Intel architectures except Nehalem, for which it
is 133.3 MHz
.
If this method fails or is unavailable, Linux appears to determine the TSC clock speed through a [calibration] lxr3 against hardware timers.
For now, we don't provide the ability to convert cycles to time.
This library is BSD-licensed.
We are happy to receive bug reports, fixes, documentation enhancements, and other improvements.
Please report bugs via the github issue tracker.
Master git repository:
git clone git://github.com/dterei/gotsc.git
This library is written and maintained by David Terei, code@davidterei.com.