alihan/fn-serverless
1
0
Fork 0
mirror of https://github.com/fnproject/fn.git synced 2022-10-28 21:29:17 +03:00
Code Issues Packages Projects Releases Wiki Activity

fn: eviction resource correction (#1282)

Browse Source 
Previously evictor did not perform an eviction
if total cpu/mem of evictable containers was less
than requested cpu/mem. With this change, we
try to perform evictions based on actual needed cpu & mem
reported by resource tracker.
This commit is contained in:
Tolga Ceylan
2018-10-25 03:10:19 -07:00
committed by Dario Domizioli
parent e8e6ed2fd0
commit bf41789af2
2 changed files with 44 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
api/agent/agent.go
View File

@@ -452,6 +452,7 @@ func (a *agent) checkLaunch(ctx context.Context, call *call, notifyChan chan err
mem := call.Memory + uint64(call.TmpFsSize)
var notifyChans []chan struct{}
var tok ResourceToken
// WARNING: Tricky flow below. We are here because: isNewContainerNeeded is true,
// in other words, we need to launch a new container at this time due to high load.
@@ -471,12 +472,28 @@ func (a *agent) checkLaunch(ctx context.Context, call *call, notifyChan chan err
// Non-blocking mode only applies to cpu+mem, and if isNewContainerNeeded decided that we do not
// need to start a new container, then waiters will wait.
select {
case tok := <-a.resources.GetResourceToken(ctx, mem, call.CPUs, isNB):
if tok != nil && tok.Error() != nil {
case tok = <-a.resources.GetResourceToken(ctx, mem, call.CPUs, isNB):
case <-time.After(a.cfg.HotPoll):
// Request routines are polling us with this a.cfg.HotPoll frequency. We can use this
// same timer to assume that we waited for cpu/mem long enough. Let's try to evict an
// idle container. We do this by submitting a non-blocking request and evicting required
// amount of resources.
select {
case tok = <-a.resources.GetResourceToken(ctx, mem, call.CPUs, true):
case <-ctx.Done(): // timeout
case <-a.shutWg.Closer(): // server shutdown
}
case <-ctx.Done(): // timeout
case <-a.shutWg.Closer(): // server shutdown
}
if tok != nil {
if tok.Error() != nil {
if tok.Error() != CapacityFull {
tryNotify(notifyChan, tok.Error())
} else {
notifyChans = a.evictor.PerformEviction(call.slotHashId, mem, uint64(call.CPUs))
needMem, needCpu := tok.NeededCapacity()
notifyChans = a.evictor.PerformEviction(call.slotHashId, needMem, uint64(needCpu))
if len(notifyChans) == 0 {
tryNotify(notifyChan, tok.Error())
}
@@ -490,17 +507,8 @@ func (a *agent) checkLaunch(ctx context.Context, call *call, notifyChan chan err
// early return (do not allow container state to switch to ContainerStateDone)
return
}
if tok != nil {
statsUtilization(ctx, a.resources.GetUtilization())
tok.Close()
}
// Request routines are polling us with this a.cfg.HotPoll frequency. We can use this
// same timer to assume that we waited for cpu/mem long enough. Let's try to evict an
// idle container.
case <-time.After(a.cfg.HotPoll):
notifyChans = a.evictor.PerformEviction(call.slotHashId, mem, uint64(call.CPUs))
case <-ctx.Done(): // timeout
case <-a.shutWg.Closer(): // server shutdown
statsUtilization(ctx, a.resources.GetUtilization())
tok.Close()
}
defer state.UpdateState(ctx, ContainerStateDone, call.slots)