go - golang程序内存不断增长，但只有一点inuse_space

Question

我有一个golang程序，它定期运行一些例程。它的内存(RSS排在首位)在某些服务器上持续增长，但在某些服务器上却没有。奇怪的是，RSS增长实例的pprof结果显示该进程仅占用了很少的inuse_space内存。

这是否意味着内存正在等待golang到GC？但是48小时后，内存从26000kb增长到74000kb，我认为golang本身应该已经做过几次GC。

我已经使用sync.Pool来避免频繁的new()ing对象。如何减少内存使用量？

详细数据附在下面，谢谢!

48小时内RSS的使用情况(以kb为单位):


inuse_space:

(pprof) top
Showing nodes accounting for 1040.73kB, 100% of 1040.73kB total
      flat  flat%   sum%        cum   cum%
  524.09kB 50.36% 50.36%   524.09kB 50.36%  sp/audit.Coll
  516.64kB 49.64%   100%   516.64kB 49.64%  sp/audit.Coll.func2
         0     0%   100%   524.09kB 50.36%  sa/core.(*pr).run
         0     0%   100%   516.64kB 49.64%  sa/fw.(*Group).Go.func1
         0     0%   100%   516.64kB 49.64%  sp/audit/dis.(*Relay).Relay
         0     0%   100%   516.64kB 49.64%  sync.(*Pool).Get
(pprof) top -cum
Showing nodes accounting for 1040.73kB, 100% of 1040.73kB total
      flat  flat%   sum%        cum   cum%
         0     0%     0%   524.09kB 50.36%  sa/core.(*pr).run
  524.09kB 50.36% 50.36%   524.09kB 50.36%  sp/audit.Coll
         0     0% 50.36%   516.64kB 49.64%  sa/fw.(*Group).Go.func1
  516.64kB 49.64%   100%   516.64kB 49.64%  sp/audit.Coll.func2
         0     0%   100%   516.64kB 49.64%  sp/audit/dis.(*Relay).Relay
         0     0%   100%   516.64kB 49.64%  sync.(*Pool).Get

alloc_space:

Showing top 10 nodes out of 181
      flat  flat%   sum%        cum   cum%
 3306.19MB 25.83% 25.83%  3306.71MB 25.84%  io.copyBuffer
 1294.76MB 10.12% 35.95%  1386.15MB 10.83%  os.(*File).readdirnames
 1121.62MB  8.76% 44.71%  1121.62MB  8.76%  syscall.anyToSockaddr
  471.56MB  3.68% 48.40%   471.56MB  3.68%  net.newFD
  413.74MB  3.23% 51.63%   413.74MB  3.23%  bytes.makeSlice
  356.52MB  2.79% 54.41%   356.52MB  2.79%  github.com/go-xorm/xorm.(*Statement).Init
  261.55MB  2.04% 56.46%   309.05MB  2.41%  os.lstatNolog
  241.03MB  1.88% 58.34%   275.03MB  2.15%  os.Readlink
  236.51MB  1.85% 60.19%   236.51MB  1.85%  reflect.packEface
  230.51MB  1.80% 61.99%   230.51MB  1.80%  net.sockaddrToUnix
(pprof) top -cum
Showing nodes accounting for 3328.69MB, 26.01% of 12798.76MB total
Dropped 439 nodes (cum <= 63.99MB)
Showing top 10 nodes out of 181
      flat  flat%   sum%        cum   cum%
    6.50MB 0.051% 0.051%  6244.66MB 48.79%  sp/filesystem.walk
         0     0% 0.051%  6244.66MB 48.79%  sp/filesystem.walkProxy
         0     0% 0.051%  6088.14MB 47.57%  sa/fw.(*WorkerPool).StartWorker.func1
         0     0% 0.051%  5073.21MB 39.64%  sa/core.(*procRuntime).run
       1MB 0.0078% 0.059%  3446.69MB 26.93%  sp/filesystem.buildFileStat
       4MB 0.031%  0.09%  3431.67MB 26.81%  sp/filesystem.Chksum
         0     0%  0.09%  3334.15MB 26.05%  io.CopyN
         0     0%  0.09%  3306.71MB 25.84%  io.Copy
 3306.19MB 25.83% 25.92%  3306.71MB 25.84%  io.copyBuffer
      11MB 0.086% 26.01%  2659.24MB 20.78%  sa/fw.(*DBO).insertKeepOld

Answer 1

从1.12版本开始，开始使用madvise的无延迟MADV_FREE模式向OS释放内存。除非系统内存不足，否则这不会在RSS中显示。要获得实际用法，您必须查看RSS-LazyFree。它在/proc中报告。