avoid quantization and use worker channel to ensure expected number of workers
This commit is contained in:
parent
f6df15cb38
commit
5f7915649d
@ -117,42 +117,68 @@ func (sds *Service) APIs() []rpc.API {
|
||||
}
|
||||
}
|
||||
|
||||
func segmentRange(workers, start, stop uint64, params sd.Params) []RangeRequest {
|
||||
segmentSize := ((stop - start) + 1) / workers
|
||||
remainder := ((stop - start) + 1) % workers
|
||||
numOfSegments := workers
|
||||
if remainder > 0 {
|
||||
numOfSegments++
|
||||
}
|
||||
segments := make([]RangeRequest, numOfSegments)
|
||||
for i := range segments {
|
||||
end := start + segmentSize - 1
|
||||
if end > stop {
|
||||
end = stop
|
||||
}
|
||||
segments[i] = RangeRequest{start, end, params}
|
||||
start = end + 1
|
||||
}
|
||||
return segments
|
||||
}
|
||||
|
||||
// Run does a one-off processing run on the provided RangeRequests + any pre-runs, exiting afterwards
|
||||
func (sds *Service) Run(rngs []RangeRequest, parallel bool) error {
|
||||
for _, preRun := range sds.preruns {
|
||||
// if the rangeSize is smaller than the number of workers
|
||||
// make sure we do synchronous processing to avoid quantization issues
|
||||
rangeSize := (preRun.Stop - preRun.Start) + 1
|
||||
numWorkers := uint64(sds.workers)
|
||||
if rangeSize < numWorkers {
|
||||
parallel = false
|
||||
}
|
||||
if parallel {
|
||||
// Chunk overall range into N subranges for workers
|
||||
chunkSize := (preRun.Stop - preRun.Start) / uint64(sds.workers)
|
||||
logrus.Infof("parallel processing prerun range (%d, %d) (%d blocks) divided into %d sized chunks with %d workers", preRun.Start, preRun.Stop,
|
||||
preRun.Stop-preRun.Start+1, chunkSize, sds.workers)
|
||||
// Sanity floor the chunk size
|
||||
if chunkSize < 100 {
|
||||
chunkSize = 100
|
||||
logrus.Infof("Computed range chunk size for each worker is too small, defaulting to 100")
|
||||
}
|
||||
rangeSize, rangeSize/numWorkers, numWorkers)
|
||||
workChan := make(chan RangeRequest)
|
||||
quitChan := make(chan struct{})
|
||||
// spin up numWorkers number of worker goroutines
|
||||
wg := new(sync.WaitGroup)
|
||||
for i := 0; i < int(sds.workers); i++ {
|
||||
blockRange := RangeRequest{
|
||||
Start: preRun.Start + uint64(i)*chunkSize,
|
||||
Stop: preRun.Start + uint64(i)*chunkSize + chunkSize - 1,
|
||||
Params: preRun.Params,
|
||||
}
|
||||
// TODO(hack) this fixes quantization
|
||||
if blockRange.Stop < preRun.Stop && preRun.Stop-blockRange.Stop < chunkSize {
|
||||
blockRange.Stop = preRun.Stop
|
||||
}
|
||||
for i := 0; i < int(numWorkers); i++ {
|
||||
wg.Add(1)
|
||||
go func(id int) {
|
||||
defer wg.Done()
|
||||
logrus.Infof("prerun worker %d processing range (%d, %d)", id, blockRange.Start, blockRange.Stop)
|
||||
for j := blockRange.Start; j <= blockRange.Stop; j++ {
|
||||
if err := sds.WriteStateDiffAt(j, blockRange.Params); err != nil {
|
||||
logrus.Errorf("error writing statediff at height %d in range (%d, %d) : %v", i, blockRange.Start, blockRange.Stop, err)
|
||||
for {
|
||||
select {
|
||||
case workerSegment := <-workChan:
|
||||
for j := workerSegment.Start; j <= workerSegment.Stop; j++ {
|
||||
if err := sds.WriteStateDiffAt(j, workerSegment.Params); err != nil {
|
||||
logrus.Errorf("error writing statediff at height %d in range (%d, %d) : %v", id, workerSegment.Start, workerSegment.Stop, err)
|
||||
}
|
||||
}
|
||||
logrus.Infof("prerun worker %d finished processing range (%d, %d)", id, workerSegment.Start, workerSegment.Stop)
|
||||
case <-quitChan:
|
||||
return
|
||||
}
|
||||
}
|
||||
logrus.Infof("prerun worker %d finished processing range (%d, %d)", id, blockRange.Start, blockRange.Stop)
|
||||
}(i)
|
||||
}
|
||||
// break range up into segments
|
||||
segments := segmentRange(numWorkers, preRun.Start, preRun.Stop, preRun.Params)
|
||||
// send the segments to the work channel
|
||||
for _, segment := range segments {
|
||||
workChan <- segment
|
||||
}
|
||||
close(quitChan)
|
||||
wg.Wait()
|
||||
} else {
|
||||
logrus.Infof("sequential processing prerun range (%d, %d)", preRun.Start, preRun.Stop)
|
||||
|
Loading…
Reference in New Issue
Block a user