Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

fix: handle CUDA errors by splitting proof requests into single-block chunks #313

Closed
wants to merge 1 commit into from
Closed

fix: handle CUDA errors by splitting proof requests into single-block chunks #313

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
69 changes: 53 additions & 16 deletions proposer/op/proposer/prove.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,26 @@ import (
)

const PROOF_STATUS_TIMEOUT = 30 * time.Second
const WITNESSGEN_TIMEOUT = 20 * time.Minute

// Status constants
const (
SP1FulfillmentStatusUnfulfillable = 2
SP1FulfillmentStatusFulfilled = 1
SP1ExecutionStatusUnexecutable = 2
)

// ProofStatusResponse represents the response from the prover service
type ProofStatusResponse struct {
FulfillmentStatus int `json:"fulfillment_status"`
ExecutionStatus int `json:"execution_status"`
Error string `json:"error"`
Proof []byte `json:"proof"`
}

// This limit is set to prevent overloading the witness generation server. Until Kona improves their native I/O API (https://github.com/anton-rs/kona/issues/553)
// the maximum number of concurrent witness generation requests is roughly num_cpu / 2. Set it to 5 for now to be safe.
const MAX_CONCURRENT_WITNESS_GEN = 5

// Process all of requests in PROVING state.
func (l *L2OutputSubmitter) ProcessProvingRequests() error {
Expand Down Expand Up @@ -70,7 +90,7 @@ func (l *L2OutputSubmitter) ProcessWitnessgenRequests() error {
for _, req := range reqs {
// If the request has been in the WITNESSGEN state for longer than the timeout, set status to FAILED.
// This is a catch-all in case the witness generation state update failed.
if req.LastUpdatedTime+uint64(l.Cfg.WitnessGenTimeout) < uint64(time.Now().Unix()) {
if req.LastUpdatedTime+uint64(WITNESSGEN_TIMEOUT.Seconds()) < uint64(time.Now().Unix()) {
// Retry the request if it timed out.
l.RetryRequest(req, ProofStatusResponse{})
}
Expand All @@ -94,18 +114,36 @@ func (l *L2OutputSubmitter) RetryRequest(req *ent.ProofRequest, status ProofStat
// This is likely caused by an SP1 OOM due to a large block range with many transactions.
// TODO: This solution can be removed once the embedded allocator is used, because then the programs
// will never OOM.
if req.Type == proofrequest.TypeSPAN && status.ExecutionStatus == SP1ExecutionStatusUnexecutable && req.EndBlock-req.StartBlock > 1 {
// Split the request into two requests.
midBlock := (req.StartBlock + req.EndBlock) / 2
err = l.db.NewEntry(req.Type, req.StartBlock, midBlock)
if err != nil {
l.Log.Error("failed to retry first half of proof request", "err", err)
return err
if req.Type == proofrequest.TypeSPAN && (status.ExecutionStatus == SP1ExecutionStatusUnexecutable || strings.Contains(status.Error, "CUDA")) && req.EndBlock-req.StartBlock > 1 {
// For CUDA errors, split into smaller chunks
var chunks []struct{ start, end uint64 }
chunkSize := uint64(1) // Start with smallest possible chunk

if strings.Contains(status.Error, "CUDA") {
// Use very small chunks for CUDA errors
for start := req.StartBlock; start < req.EndBlock; start += chunkSize {
end := start + chunkSize
if end > req.EndBlock {
end = req.EndBlock
}
chunks = append(chunks, struct{ start, end uint64 }{start, end})
}
} else {
// Original split in half behavior
midBlock := (req.StartBlock + req.EndBlock) / 2
chunks = []struct{ start, end uint64 }{
{req.StartBlock, midBlock},
{midBlock + 1, req.EndBlock},
}
}
err = l.db.NewEntry(req.Type, midBlock, req.EndBlock)
if err != nil {
l.Log.Error("failed to retry second half of proof request", "err", err)
return err

// Create new entries for each chunk
for _, chunk := range chunks {
err = l.db.NewEntry(req.Type, chunk.start, chunk.end)
if err != nil {
l.Log.Error("failed to create new chunk entry", "start", chunk.start, "end", chunk.end, "err", err)
return err
}
}
} else {
// Retry the same request.
Expand Down Expand Up @@ -157,7 +195,7 @@ func (l *L2OutputSubmitter) RequestQueuedProofs(ctx context.Context) error {

// The number of witness generation requests is capped at MAX_CONCURRENT_WITNESS_GEN. This prevents overloading the machine with processes spawned by the witness generation server.
// Once https://github.com/anton-rs/kona/issues/553 is fixed, we may be able to remove this check.
if witnessGenProofs >= int(l.Cfg.MaxConcurrentWitnessGen) {
if witnessGenProofs >= MAX_CONCURRENT_WITNESS_GEN {
l.Log.Info("max witness generation reached, waiting for next cycle")
return nil
}
Expand Down Expand Up @@ -326,14 +364,13 @@ func (l *L2OutputSubmitter) makeProofRequest(proofType proofrequest.Type, jsonBo
}
req.Header.Set("Content-Type", "application/json")

timeout := time.Duration(l.Cfg.WitnessGenTimeout) * time.Second
client := &http.Client{Timeout: timeout}
client := &http.Client{Timeout: WITNESSGEN_TIMEOUT}
resp, err := client.Do(req)
if err != nil {
if netErr, ok := err.(net.Error); ok && netErr.Timeout() {
l.Log.Error("Witness generation request timed out", "err", err)
l.Metr.RecordWitnessGenFailure("Timeout")
return nil, fmt.Errorf("request timed out after %s: %w", timeout, err)
return nil, fmt.Errorf("request timed out after %s: %w", WITNESSGEN_TIMEOUT, err)
}
return nil, fmt.Errorf("failed to send request: %w", err)
}
Expand Down