93e4656a27
Before this change workers can only be allocated one GPU task, regardless of how much of the GPU resources that task uses, or how many GPUs are in the system. This makes GPUUtilization a float which can represent that a task needs a portion, or multiple GPUs. GPUs are accounted for like RAM and CPUs so that workers with more GPUs can be allocated more tasks. A known issue is that PC2 cannot use multiple GPUs. And even if the worker has multiple GPUs and is allocated multiple PC2 tasks, those tasks will only run on the first GPU. This could result in unexpected behavior when a worker with multiple GPUs is assigned multiple PC2 tasks. But this should not suprise any existing users who upgrade, as any existing users who run workers with multiple GPUs should already know this and be running a worker per GPU for PC2. But now those users have the freedom to customize the GPU utilization of PC2 to be less than one and effectively run multiple PC2 processes in a single worker. C2 is capable of utilizing multiple GPUs, and now workers can be customized for C2 accordingly.
143 lines
3.7 KiB
Go
143 lines
3.7 KiB
Go
package sectorstorage
|
|
|
|
import (
|
|
"sync"
|
|
|
|
"github.com/filecoin-project/lotus/extern/sector-storage/storiface"
|
|
)
|
|
|
|
func (a *activeResources) withResources(id WorkerID, wr storiface.WorkerInfo, r Resources, locker sync.Locker, cb func() error) error {
|
|
for !a.canHandleRequest(r, id, "withResources", wr) {
|
|
if a.cond == nil {
|
|
a.cond = sync.NewCond(locker)
|
|
}
|
|
a.waiting++
|
|
a.cond.Wait()
|
|
a.waiting--
|
|
}
|
|
|
|
a.add(wr.Resources, r)
|
|
|
|
err := cb()
|
|
|
|
a.free(wr.Resources, r)
|
|
|
|
return err
|
|
}
|
|
|
|
// must be called with the same lock as the one passed to withResources
|
|
func (a *activeResources) hasWorkWaiting() bool {
|
|
return a.waiting > 0
|
|
}
|
|
|
|
func (a *activeResources) add(wr storiface.WorkerResources, r Resources) {
|
|
if r.GPUUtilization > 0 {
|
|
a.gpuUsed += r.GPUUtilization
|
|
}
|
|
a.cpuUse += r.Threads(wr.CPUs)
|
|
a.memUsedMin += r.MinMemory
|
|
a.memUsedMax += r.MaxMemory
|
|
}
|
|
|
|
func (a *activeResources) free(wr storiface.WorkerResources, r Resources) {
|
|
if r.GPUUtilization > 0 {
|
|
a.gpuUsed -= r.GPUUtilization
|
|
}
|
|
a.cpuUse -= r.Threads(wr.CPUs)
|
|
a.memUsedMin -= r.MinMemory
|
|
a.memUsedMax -= r.MaxMemory
|
|
|
|
if a.cond != nil {
|
|
a.cond.Broadcast()
|
|
}
|
|
}
|
|
|
|
// canHandleRequest evaluates if the worker has enough available resources to
|
|
// handle the request.
|
|
func (a *activeResources) canHandleRequest(needRes Resources, wid WorkerID, caller string, info storiface.WorkerInfo) bool {
|
|
if info.IgnoreResources {
|
|
// shortcircuit; if this worker is ignoring resources, it can always handle the request.
|
|
return true
|
|
}
|
|
|
|
res := info.Resources
|
|
|
|
// TODO: dedupe needRes.BaseMinMemory per task type (don't add if that task is already running)
|
|
memNeeded := needRes.MinMemory + needRes.BaseMinMemory
|
|
memUsed := a.memUsedMin
|
|
// assume that MemUsed can be swapped, so only check it in the vmem Check
|
|
memAvail := res.MemPhysical - memUsed
|
|
if memNeeded > memAvail {
|
|
log.Debugf("sched: not scheduling on worker %s for %s; not enough physical memory - need: %dM, have %dM available", wid, caller, memNeeded/mib, memAvail/mib)
|
|
return false
|
|
}
|
|
|
|
vmemNeeded := needRes.MaxMemory + needRes.BaseMinMemory
|
|
vmemUsed := a.memUsedMax
|
|
if vmemUsed < res.MemUsed+res.MemSwapUsed {
|
|
vmemUsed = res.MemUsed + res.MemSwapUsed
|
|
}
|
|
vmemAvail := res.MemPhysical + res.MemSwap - vmemUsed
|
|
|
|
if vmemNeeded > vmemAvail {
|
|
log.Debugf("sched: not scheduling on worker %s for %s; not enough virtual memory - need: %dM, have %dM available", wid, caller, vmemNeeded/mib, vmemAvail/mib)
|
|
return false
|
|
}
|
|
|
|
if a.cpuUse+needRes.Threads(res.CPUs) > res.CPUs {
|
|
log.Debugf("sched: not scheduling on worker %s for %s; not enough threads, need %d, %d in use, target %d", wid, caller, needRes.Threads(res.CPUs), a.cpuUse, res.CPUs)
|
|
return false
|
|
}
|
|
|
|
if len(res.GPUs) > 0 && needRes.GPUUtilization > 0 {
|
|
if a.gpuUsed+needRes.GPUUtilization > float64(len(res.GPUs)) {
|
|
log.Debugf("sched: not scheduling on worker %s for %s; GPU(s) in use", wid, caller)
|
|
return false
|
|
}
|
|
}
|
|
|
|
return true
|
|
}
|
|
|
|
func (a *activeResources) utilization(wr storiface.WorkerResources) float64 {
|
|
var max float64
|
|
|
|
cpu := float64(a.cpuUse) / float64(wr.CPUs)
|
|
max = cpu
|
|
|
|
memUsed := a.memUsedMin
|
|
if memUsed < wr.MemUsed {
|
|
memUsed = wr.MemUsed
|
|
}
|
|
memMin := float64(memUsed) / float64(wr.MemPhysical)
|
|
if memMin > max {
|
|
max = memMin
|
|
}
|
|
|
|
vmemUsed := a.memUsedMax
|
|
if a.memUsedMax < wr.MemUsed+wr.MemSwapUsed {
|
|
vmemUsed = wr.MemUsed + wr.MemSwapUsed
|
|
}
|
|
memMax := float64(vmemUsed) / float64(wr.MemPhysical+wr.MemSwap)
|
|
|
|
if memMax > max {
|
|
max = memMax
|
|
}
|
|
|
|
return max
|
|
}
|
|
|
|
func (wh *workerHandle) utilization() float64 {
|
|
wh.lk.Lock()
|
|
u := wh.active.utilization(wh.info.Resources)
|
|
u += wh.preparing.utilization(wh.info.Resources)
|
|
wh.lk.Unlock()
|
|
wh.wndLk.Lock()
|
|
for _, window := range wh.activeWindows {
|
|
u += window.allocated.utilization(wh.info.Resources)
|
|
}
|
|
wh.wndLk.Unlock()
|
|
|
|
return u
|
|
}
|