Those nanosleep calls are at 10ms intervals - i.e. 100 calls per second. I think this is normal behaviour of the Go scheduler, which signals itself every 10ms to force a preemptive context switch. That feature was introduced in go 1.14, at which point you were able to disable it at runtime setting environment variable GODEBUG=asyncpreemptoff=1 (I haven't tested whether you can still do this).
There were some problems when this feature was first introduced (see this <https://github.com/kubernetes/kubernetes/issues/92521> and this <https://github.com/cockroachdb/cockroach/pull/52348>) but improvements were made in 1.15 and 1.16. I don't see anything in the 1.17 release notes <https://go.dev/doc/go1.17> which would affect this. But in any case: 1. That won't cause CPU to go to 100% 2. I can't think why there should be any difference between "go run", versus "go build" and running the resulting executable. 3. I can't think why changing HTTP timeouts from 30 seconds to 1 second should make any difference. Point (2) is going to make the problem hard to reproduce for other people. To start with, I think you need to be more specific about: - exactly what versions of go you're comparing ("go version"), and where you got them from (e.g. directly from go downloads page or from OS vendor packages) - what exact Linux kernel you're running (uname -a) - what exact version and distro of Linux you're using - your CPU type (grep "model" /proc/cpuinfo | sort | uniq -c) - anything that's potentially different about your environment - e.g. are you running Linux under an emulation layer like WSL (Windows) On Sunday, 10 April 2022 at 23:02:40 UTC+1 scorr...@gmail.com wrote: > Hi, > > On Linux, if I compile and run this simple program: > > package main > > import ( > "fmt" > "net/http" > "time" > ) > > type x struct{} > > func (x) ServeHTTP(w http.ResponseWriter, r *http.Request) { > w.Write([]byte("aaaa")) > } > > func main() { > sp := &http.Server{ > ReadHeaderTimeout: 5 * time.Second, > ReadTimeout: 5 * time.Second, > WriteTimeout: 5 * time.Second, > IdleTimeout: 30 * time.Second, > Addr: ":8080", > Handler: x{}, > } > > fmt.Println(sp.ListenAndServe()) > } > > After a while I start to see constant CPU activity (1% to 7%) even when > the server is idle. > > If I run "sudo strace -f -c -p <pid>" I get constant very fast calls like: > > [pid 131277] 23:42:16 nanosleep({tv_sec=0, tv_nsec=10000000}, NULL) = 0 > [pid 131277] 23:42:16 nanosleep({tv_sec=0, tv_nsec=10000000}, NULL) = 0 > [pid 131277] 23:42:16 nanosleep({tv_sec=0, tv_nsec=10000000}, NULL) = 0 > [pid 131277] 23:42:16 nanosleep({tv_sec=0, tv_nsec=10000000}, NULL) = 0 > [pid 131277] 23:42:16 nanosleep({tv_sec=0, tv_nsec=10000000}, NULL) = 0 > .... > > It never stops. Sometimes the CPU goes to 100% for a while and then go > back down to 1%.. This happens only in one or two cores. > > This doesn't happen with Go 1.16. I have been checking all versions and it > starts in Go 1.17 > > If I run the program with go run instead of compiling and running the > binary, I can't reproduce it. > > Sometimes it starts after a couple of http calls and sometimes I do a few > "wrk -d2 http://localhost:8080"; calls to trigger it (wrk is a web stress > tool). But as soon as it starts making the nanosleep calls, it doesn't > stop, no matter how long the server is iddle. > > If I remove the timeout values or set them to a short period, like one > second for example, I am also not able to reproduce it. But with longer > values it happens very quickly. I am surprised of not seing it reported. > > Do you know what can be going on? > > Thank you! > > -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/golang-nuts/7b4b7f21-8587-4936-bc29-d5cc58206d18n%40googlegroups.com.
