Update all vendoring to current releases. Signed-off-by: Ben Kochie <superq@gmail.com>
5.4 KiB
Perf
This package is a go library for interacting with the perf
subsystem in
Linux. I had trouble finding a golang perf library so I decided to write this
by using the linux's perf as a reference. This library allows you to do things
like see how many CPU instructions a function takes (roughly), profile a
process for various hardware events, and other interesting things. Note that
because the go scheduler can schedule a goroutine across many OS threads it
becomes rather difficult to get an exact profile of an invididual goroutine.
However, a few tricks can be used; first a call to
[runtime.LockOSThread](https://golang.org/pkg/runtime/#LockOSThread)
to lock
the current goroutine to an OS thread. Second a call to
[unix.SchedSetaffinity](https://godoc.org/golang.org/x/sys/unix#SchedSetaffinity)
,
with a CPU set mask set. Note that if the pid argument is set 0 the calling
thread is used (the thread that was just locked). Before using this library you
should probably read the
perf_event_open
man page which this library uses heavily.
Use Cases
If you are looking to interact with the perf subsystem directly with
perf_event_open
syscall than this library is most likely for you. A large
number of the utility methods in this package should only be used for testing
and/or debugging performance issues. This is due to the nature of the go
runtime being extremely tricky to profile on the goroutine level, with the
exception of a long running worker goroutine locked to an OS thread. Eventually
this library could be used to implement many of the features of perf
but in
pure Go. Currently this library is used in
node_exporter as well as
perf_exporter, which is a
Prometheus exporter for perf related metrics.
Caveats
- Some utility functions will call
runtime.LockOSThread
for you, they will also unlock the thread after profiling. Note using these utility functions will incur significant overhead (~4ms). - Overflow handling is not implemented.
Setup
Most likely you will need to tweak some system settings unless you are running
as root. From man perf_event_open
:
perf_event related configuration files
Files in /proc/sys/kernel/
/proc/sys/kernel/perf_event_paranoid
The perf_event_paranoid file can be set to restrict access to the performance counters.
2 allow only user-space measurements (default since Linux 4.6).
1 allow both kernel and user measurements (default before Linux 4.6).
0 allow access to CPU-specific data but not raw tracepoint samples.
-1 no restrictions.
The existence of the perf_event_paranoid file is the official method for determining if a kernel supports perf_event_open().
/proc/sys/kernel/perf_event_max_sample_rate
This sets the maximum sample rate. Setting this too high can allow users to sample at a rate that impacts overall machine performance and potentially lock up the machine. The default value is 100000 (samples per
second).
/proc/sys/kernel/perf_event_max_stack
This file sets the maximum depth of stack frame entries reported when generating a call trace.
/proc/sys/kernel/perf_event_mlock_kb
Maximum number of pages an unprivileged user can mlock(2). The default is 516 (kB).
Example
Say you wanted to see how many CPU instructions a particular function took:
package main
import (
"fmt"
"log"
"github.com/hodgesds/perf-utils"
)
func foo() error {
var total int
for i:=0;i<1000;i++ {
total++
}
return nil
}
func main() {
profileValue, err := perf.CPUInstructions(foo)
if err != nil {
log.Fatal(err)
}
fmt.Printf("CPU instructions: %+v\n", profileValue)
}
Benchmarks
To profile a single function call there is an overhead of ~0.4ms.
$ go test -bench=BenchmarkCPUCycles .
goos: linux
goarch: amd64
pkg: github.com/hodgesds/perf-utils
BenchmarkCPUCycles-8 3000 397924 ns/op 32 B/op 1 allocs/op
PASS
ok github.com/hodgesds/perf-utils 1.255s
The Profiler
interface has low overhead and suitable for many use cases:
$ go test -bench=BenchmarkProfiler .
goos: linux
goarch: amd64
pkg: github.com/hodgesds/perf-utils
BenchmarkProfiler-8 3000000 488 ns/op 32 B/op 1 allocs/op
PASS
ok github.com/hodgesds/perf-utils 1.981s
BPF Support
BPF is supported by using the BPFProfiler
which is available via the
ProfileTracepoint
function. To use BPF you need to create the BPF program and
then call AttachBPF
with the file descriptor of the BPF program.
Misc
Originally I set out to use go generate
to build Go structs that were
compatible with perf, I found a really good
article
on how to do so. Eventually, after digging through some of the /x/sys/unix
code I found pretty much what I was needed. However, I think if you are
interested in interacting with the kernel it is a worthwhile read.