Open
Conversation
…g gotoni binary on each instance using scp
…ancer.go to use sglang_metrics.go for updating availability state, remove WrapHandler to make Handler() flow more streamlined, update lb.go, and remove claude bloat.
…uest into trie before execution [ ] network policy can use the existing metric polling with timestamps to get and update real time inter-node latency metrics
… TestClusterSGLang test
…istral-7b for non-gated model serving
…er with go->rust bindings. Integrated GetTokenCount into new GORGO Policy for loadbalancer.go. Added FlushSGLangCache() in sglang_metrics.go for use in accurate logistic regression of server response times varying with token count from 100 requests in wildchat_flat.jsonl
…r tokenizer_cgo.go compilation on Linux, and include tokenizer test directly in TestClusterSGLang
… when running TestClusterSGLang. Added endpoint in loadbalancer.go for testing tokenizer functionality
…ter_test.go mods and new testing script.
…glang metrics and loadbalancer event logs
…rest cluster endpoint.
…the load balancer endpoint for metrics
…and updated project structure
…o76/gotoni into benchmark-load-balancing
…ancer/delete binary before reuploading gotoni binary
…o76/gotoni into benchmark-load-balancing Merge changes into benchmark-load-balancing
…o76/gotoni into benchmark-load-balancing hh
…o76/gotoni into benchmark-load-balancing
…o76/gotoni into benchmark-load-balancing dfkj.
There was a problem hiding this comment.
Pull request overview
This PR introduces a comprehensive load balancing infrastructure for distributed GPU cluster management. It adds event-driven tracing (Perfetto-compatible), tokenizer sidecar integration (Rust), SGLang cluster management, and multiple load balancing policies (GORGO, prefix-tree, least-loaded). The PR also refactors instance state management to use Lambda API as the source of truth instead of local database storage.
Changes:
- Added distributed tracing system with Perfetto format support
- Implemented Rust tokenizer sidecar with Unix socket communication
- Created cluster management infrastructure for SGLang deployments
- Added load balancing policies and configuration system
- Removed local DB dependency for instance state, using Lambda API instead
- Added comprehensive benchmarking tools using WildChat-1M dataset
Reviewed changes
Copilot reviewed 39 out of 42 changed files in this pull request and generated 12 comments.
Show a summary per file
| File | Description |
|---|---|
| pkg/trace/trace.go | New event-driven tracing system (Perfetto format) |
| pkg/cluster/trace.go | Duplicate of pkg/trace/trace.go |
| pkg/tokenizer/*.go | Tokenizer sidecar management and client |
| pkg/tokenizer/main.rs | Rust tokenizer service with Unix socket |
| pkg/sglang/sglang.go | SGLang Docker deployment and diagnostics |
| pkg/cluster/*.go | Cluster management, metrics, build/deploy |
| pkg/config/constants.go | Centralized configuration constants |
| pkg/remote/client.go | Refactored SSH key lookup from Lambda API |
| cmd/*.go | New commands for cluster, tokenizer, LB management |
| benchmark/*.py | Geo-distributed benchmarking tools |
| go.mod | Added tokenizers dependency (with conflicts) |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+52
to
+115
| .expect("Failed to download tokenizer.json"); | ||
|
|
||
| if !response.status().is_success() { | ||
| panic!("Failed to download tokenizer: HTTP {}", response.status()); | ||
| } | ||
|
|
||
| let bytes = response.bytes().await | ||
| .expect("Failed to read tokenizer bytes"); | ||
|
|
||
| // Cache the tokenizer | ||
| std::fs::write(TOKENIZER_CACHE_PATH, &bytes) | ||
| .expect("Failed to cache tokenizer"); | ||
|
|
||
| println!("Tokenizer cached to: {}", TOKENIZER_CACHE_PATH); | ||
|
|
||
| Tokenizer::from_bytes(&bytes) | ||
| .expect("Failed to parse tokenizer.json") | ||
| } | ||
|
|
||
| #[tokio::main] | ||
| async fn main() { | ||
| println!("Loading tokenizer model: {}", MODEL_ID); | ||
|
|
||
| // Load the Mistral tokenizer | ||
| let tokenizer: Arc<Tokenizer> = Arc::new(load_or_download_tokenizer().await); | ||
| println!("Tokenizer loaded successfully"); | ||
|
|
||
| // Build the router with shared state | ||
| let tk_for_count = Arc::clone(&tokenizer); | ||
|
|
||
| let app = Router::new() | ||
| .route("/count", post(move |Json(payload): Json<TokenizeRequest>| { | ||
| let tk: Arc<Tokenizer> = Arc::clone(&tk_for_count); | ||
| async move { | ||
| let encoding = tk.encode(payload.text, false).expect("Tokenization failed"); | ||
| Json(TokenizeResponse { | ||
| count: encoding.get_ids().len(), | ||
| }) | ||
| } | ||
| })) | ||
| .route("/health", get(|| async { | ||
| Json(HealthResponse { | ||
| status: "ok".to_string(), | ||
| model: MODEL_ID.to_string(), | ||
| }) | ||
| })); | ||
|
|
||
| // Unix Domain Socket path | ||
| let socket_path = "/tmp/tokenizer.sock"; | ||
|
|
||
| // Clean up old socket if it exists | ||
| let _ = std::fs::remove_file(socket_path); | ||
|
|
||
| println!("Starting tokenizer sidecar on {}", socket_path); | ||
|
|
||
| let listener = UnixListener::bind(socket_path).expect("Failed to bind to socket"); | ||
|
|
||
| // Make socket world-readable/writable | ||
| std::fs::set_permissions(socket_path, std::fs::Permissions::from_mode(0o777)) | ||
| .expect("Failed to set socket permissions"); | ||
|
|
||
| // Serve using hyper with Unix socket | ||
| loop { | ||
| let (stream, _) = listener.accept().await.expect("Accept failed"); |
There was a problem hiding this comment.
The tokenizer sidecar uses expect() for error handling in several places (lines 52, 59, 63, 68, 86, 107, 111, 115), which will panic the entire process if the operation fails. For a production service, these should use proper error handling with ? operator or unwrap_or_else() with logging to allow graceful degradation rather than crashing.
Comment on lines
+493
to
501
| // RemoveInstanceFromConfig removes an instance from the config | ||
| // Deprecated: Instance state is now managed by Lambda API, not local DB. | ||
| // This function is kept for backwards compatibility but does nothing meaningful. | ||
| func RemoveInstanceFromConfig(instanceID string) error { | ||
| // Instance state is managed by Lambda API, not local DB | ||
| // No action needed - when an instance is terminated via Lambda API, | ||
| // it will no longer appear in ListRunningInstances | ||
| return nil | ||
| } |
There was a problem hiding this comment.
The RemoveInstanceFromConfig function is deprecated but returns no error and does nothing. Callers expecting this function to clean up instance state will silently fail. Consider either removing this function entirely or returning an error indicating it's no longer supported to make the deprecation explicit.
Comment on lines
+93
to
+106
| tokenizerMu.RLock() | ||
| available := tokenizerAvailable | ||
| tokenizerMu.RUnlock() | ||
|
|
||
| if !available { | ||
| // Try to reconnect periodically | ||
| go func() { | ||
| tokenizerMu.Lock() | ||
| defer tokenizerMu.Unlock() | ||
| if !tokenizerAvailable { | ||
| tokenizerAvailable = checkTokenizerHealth() | ||
| } | ||
| }() | ||
| return fallbackTokenCount(text) |
There was a problem hiding this comment.
The pkg/cluster/tokenizer.go file has a potential race condition. The tokenizerAvailable flag is read without lock on line 94 and checked again on line 102-103, but another goroutine could modify it between these checks. The reconnection goroutine spawned on line 99 should be synchronized to prevent multiple concurrent reconnection attempts.
Comment on lines
+1
to
+573
| /* | ||
| Copyright © 2025 ALESSIO TONIOLO | ||
|
|
||
| trace.go provides event-driven tracing for the load balancer. | ||
|
|
||
| Design principles (inspired by PyTorch profiler): | ||
| - NOT constantly running - only active when explicitly started | ||
| - Event-driven with hooks inserted at key points | ||
| - Produces Perfetto-compatible traces (Chrome Trace Format) | ||
| - Supports distributed tracing across nodes via header propagation | ||
| - Minimal overhead when disabled | ||
|
|
||
| Usage: | ||
| 1. Start trace via endpoint: POST /lb/trace/start | ||
| 2. Run workload (benchmark) | ||
| 3. Stop trace via endpoint: POST /lb/trace/stop | ||
| 4. Returns Perfetto JSON that can be loaded in chrome://tracing or ui.perfetto.dev | ||
| */ | ||
| package serve | ||
|
|
||
| import ( | ||
| "encoding/json" | ||
| "fmt" | ||
| "sync" | ||
| "sync/atomic" | ||
| "time" | ||
| ) | ||
|
|
||
| // TraceEvent represents a single event in Chrome Trace Format (Perfetto-compatible) | ||
| // See: https://docs.google.com/document/d/1CvAClvFfyA5R-PhYUmn5OOQtYMH4h6I0nSsKchNAySU | ||
| type TraceEvent struct { | ||
| Name string `json:"name"` // Event name | ||
| Cat string `json:"cat"` // Category (e.g., "lb", "forward", "queue") | ||
| Ph string `json:"ph"` // Phase: B=begin, E=end, X=complete, i=instant | ||
| Ts int64 `json:"ts"` // Timestamp in microseconds | ||
| Dur int64 `json:"dur,omitempty"` // Duration in microseconds (for ph=X) | ||
| Pid int `json:"pid"` // Process ID (node ID hash) | ||
| Tid int `json:"tid"` // Thread ID (request ID hash) | ||
| Args map[string]interface{} `json:"args,omitempty"` // Additional metadata | ||
| ID string `json:"id,omitempty"` // For async events (flow) | ||
| Scope string `json:"s,omitempty"` // Scope for flow events | ||
| BindingPt string `json:"bp,omitempty"` // Binding point for flow events | ||
| } | ||
|
|
||
| // TraceSession manages an active tracing session | ||
| type TraceSession struct { | ||
| ID string | ||
| StartTime time.Time | ||
| NodeID string | ||
|
|
||
| events []TraceEvent | ||
| eventsMu sync.Mutex | ||
|
|
||
| // Atomic flag for fast "is tracing active" check | ||
| active atomic.Bool | ||
|
|
||
| // For distributed tracing - events received from other nodes | ||
| remoteEvents []TraceEvent | ||
| remoteEventsMu sync.Mutex | ||
| } | ||
|
|
||
| // Tracer is the global tracer instance | ||
| type Tracer struct { | ||
| session *TraceSession | ||
| sessionMu sync.RWMutex | ||
| nodeID string | ||
| nodePID int // Hash of node ID for Perfetto PID | ||
| } | ||
|
|
||
| // NewTracer creates a new tracer instance | ||
| func NewTracer(nodeID string) *Tracer { | ||
| return &Tracer{ | ||
| nodeID: nodeID, | ||
| nodePID: hashString(nodeID) % 10000, // Keep PID small for readability | ||
| } | ||
| } | ||
|
|
||
| // IsActive returns true if tracing is currently active (fast path) | ||
| func (t *Tracer) IsActive() bool { | ||
| t.sessionMu.RLock() | ||
| defer t.sessionMu.RUnlock() | ||
| return t.session != nil && t.session.active.Load() | ||
| } | ||
|
|
||
| // Start begins a new tracing session | ||
| func (t *Tracer) Start() string { | ||
| t.sessionMu.Lock() | ||
| defer t.sessionMu.Unlock() | ||
|
|
||
| sessionID := fmt.Sprintf("trace-%d", time.Now().UnixNano()) | ||
| t.session = &TraceSession{ | ||
| ID: sessionID, | ||
| StartTime: time.Now(), | ||
| NodeID: t.nodeID, | ||
| events: make([]TraceEvent, 0, 1000), | ||
| } | ||
| t.session.active.Store(true) | ||
|
|
||
| // Record trace start event | ||
| t.session.events = append(t.session.events, TraceEvent{ | ||
| Name: "trace_session", | ||
| Cat: "trace", | ||
| Ph: "B", | ||
| Ts: 0, // Relative to trace start | ||
| Pid: t.nodePID, | ||
| Tid: 0, | ||
| Args: map[string]interface{}{ | ||
| "session_id": sessionID, | ||
| "node_id": t.nodeID, | ||
| }, | ||
| }) | ||
|
|
||
| return sessionID | ||
| } | ||
|
|
||
| // Stop ends the tracing session and returns the Perfetto JSON | ||
| func (t *Tracer) Stop() ([]byte, error) { | ||
| t.sessionMu.Lock() | ||
| defer t.sessionMu.Unlock() | ||
|
|
||
| if t.session == nil { | ||
| return nil, fmt.Errorf("no active trace session") | ||
| } | ||
|
|
||
| t.session.active.Store(false) | ||
|
|
||
| // Record trace end event | ||
| endTs := time.Since(t.session.StartTime).Microseconds() | ||
| t.session.events = append(t.session.events, TraceEvent{ | ||
| Name: "trace_session", | ||
| Cat: "trace", | ||
| Ph: "E", | ||
| Ts: endTs, | ||
| Pid: t.nodePID, | ||
| Tid: 0, | ||
| }) | ||
|
|
||
| // Merge remote events | ||
| t.session.eventsMu.Lock() | ||
| t.session.remoteEventsMu.Lock() | ||
| allEvents := make([]TraceEvent, 0, len(t.session.events)+len(t.session.remoteEvents)) | ||
| allEvents = append(allEvents, t.session.events...) | ||
| allEvents = append(allEvents, t.session.remoteEvents...) | ||
| t.session.remoteEventsMu.Unlock() | ||
| t.session.eventsMu.Unlock() | ||
|
|
||
| // Build Perfetto trace - just the events array for direct compatibility | ||
| // Perfetto/Chrome Trace Viewer expects a plain array or {"traceEvents": [...]} | ||
| // We output just the array for maximum compatibility | ||
|
|
||
| // Add metadata as a special "M" (metadata) event at the start | ||
| metadataEvent := TraceEvent{ | ||
| Name: "process_name", | ||
| Cat: "__metadata", | ||
| Ph: "M", | ||
| Ts: 0, | ||
| Pid: t.nodePID, | ||
| Tid: 0, | ||
| Args: map[string]interface{}{ | ||
| "name": t.nodeID, | ||
| }, | ||
| } | ||
|
|
||
| // Prepend metadata event | ||
| finalEvents := make([]TraceEvent, 0, len(allEvents)+1) | ||
| finalEvents = append(finalEvents, metadataEvent) | ||
| finalEvents = append(finalEvents, allEvents...) | ||
|
|
||
| // Clear session | ||
| t.session = nil | ||
|
|
||
| return json.MarshalIndent(finalEvents, "", " ") | ||
| } | ||
|
|
||
| // GetStatus returns current trace status | ||
| func (t *Tracer) GetStatus() map[string]interface{} { | ||
| t.sessionMu.RLock() | ||
| defer t.sessionMu.RUnlock() | ||
|
|
||
| if t.session == nil { | ||
| return map[string]interface{}{ | ||
| "active": false, | ||
| "node_id": t.nodeID, | ||
| } | ||
| } | ||
|
|
||
| t.session.eventsMu.Lock() | ||
| eventCount := len(t.session.events) | ||
| t.session.eventsMu.Unlock() | ||
|
|
||
| return map[string]interface{}{ | ||
| "active": t.session.active.Load(), | ||
| "session_id": t.session.ID, | ||
| "node_id": t.nodeID, | ||
| "duration_ms": float64(time.Since(t.session.StartTime).Microseconds()) / 1000.0, | ||
| "event_count": eventCount, | ||
| "start_time": t.session.StartTime.Format(time.RFC3339), | ||
| } | ||
| } | ||
|
|
||
| // ===================================================== | ||
| // TRACE HOOKS - Insert these at key points in code | ||
| // ===================================================== | ||
|
|
||
| // TraceRequestReceived records when a request arrives at the LB | ||
| func (t *Tracer) TraceRequestReceived(requestID string, method, path, remoteAddr string) { | ||
| if !t.IsActive() { | ||
| return | ||
| } | ||
|
|
||
| t.recordEvent(TraceEvent{ | ||
| Name: "request", | ||
| Cat: "lb", | ||
| Ph: "B", // Begin | ||
| Tid: hashString(requestID) % 1000, | ||
| Args: map[string]interface{}{ | ||
| "request_id": requestID, | ||
| "method": method, | ||
| "path": path, | ||
| "remote_addr": remoteAddr, | ||
| }, | ||
| }) | ||
| } | ||
|
|
||
| // TraceRequestComplete records when a request finishes | ||
| func (t *Tracer) TraceRequestComplete(requestID string, statusCode int, servedBy string) { | ||
| if !t.IsActive() { | ||
| return | ||
| } | ||
|
|
||
| t.recordEvent(TraceEvent{ | ||
| Name: "request", | ||
| Cat: "lb", | ||
| Ph: "E", // End | ||
| Tid: hashString(requestID) % 1000, | ||
| Args: map[string]interface{}{ | ||
| "request_id": requestID, | ||
| "status_code": statusCode, | ||
| "served_by": servedBy, | ||
| }, | ||
| }) | ||
| } | ||
|
|
||
| // TraceQueueEnter records when a request enters the queue | ||
| func (t *Tracer) TraceQueueEnter(requestID string, queueDepth int) { | ||
| if !t.IsActive() { | ||
| return | ||
| } | ||
|
|
||
| t.recordEvent(TraceEvent{ | ||
| Name: "queue", | ||
| Cat: "queue", | ||
| Ph: "B", | ||
| Tid: hashString(requestID) % 1000, | ||
| Args: map[string]interface{}{ | ||
| "request_id": requestID, | ||
| "queue_depth": queueDepth, | ||
| "action": "enqueue", | ||
| }, | ||
| }) | ||
| } | ||
|
|
||
| // TraceQueueExit records when a request leaves the queue | ||
| func (t *Tracer) TraceQueueExit(requestID string, waitTimeMs float64, exitReason string) { | ||
| if !t.IsActive() { | ||
| return | ||
| } | ||
|
|
||
| t.recordEvent(TraceEvent{ | ||
| Name: "queue", | ||
| Cat: "queue", | ||
| Ph: "E", | ||
| Tid: hashString(requestID) % 1000, | ||
| Args: map[string]interface{}{ | ||
| "request_id": requestID, | ||
| "wait_time_ms": waitTimeMs, | ||
| "exit_reason": exitReason, // "local_process", "forwarded_to_peer", "timeout", "cancelled" | ||
| }, | ||
| }) | ||
| } | ||
|
|
||
| // TraceQueueForwardAttempt records when the queue processor attempts to forward a request | ||
| func (t *Tracer) TraceQueueForwardAttempt(requestID string, queueDepth int, peerSelected bool, targetPeer string) { | ||
| if !t.IsActive() { | ||
| return | ||
| } | ||
|
|
||
| t.recordEvent(TraceEvent{ | ||
| Name: "queue_forward_attempt", | ||
| Cat: "queue", | ||
| Ph: "i", // Instant event | ||
| Tid: hashString(requestID) % 1000, | ||
| Args: map[string]interface{}{ | ||
| "request_id": requestID, | ||
| "queue_depth": queueDepth, | ||
| "peer_selected": peerSelected, | ||
| "target_peer": targetPeer, | ||
| }, | ||
| }) | ||
| } | ||
|
|
||
| // TraceQueueProbe records when the queue probe runs | ||
| func (t *Tracer) TraceQueueProbe(queueDepth int, peersAvailable int, forwarded int) { | ||
| if !t.IsActive() { | ||
| return | ||
| } | ||
|
|
||
| t.recordEvent(TraceEvent{ | ||
| Name: "queue_probe", | ||
| Cat: "queue", | ||
| Ph: "i", // Instant event | ||
| Tid: 0, // Use thread 0 for system events | ||
| Args: map[string]interface{}{ | ||
| "queue_depth": queueDepth, | ||
| "peers_available": peersAvailable, | ||
| "forwarded": forwarded, | ||
| }, | ||
| }) | ||
| } | ||
|
|
||
| // TraceForwardStart records when we start forwarding to a peer | ||
| // Returns a flow ID for stitching with remote trace | ||
| func (t *Tracer) TraceForwardStart(requestID, targetNode, targetIP string) string { | ||
| if !t.IsActive() { | ||
| return "" | ||
| } | ||
|
|
||
| flowID := fmt.Sprintf("%s-%s", requestID, targetNode) | ||
|
|
||
| // Record the forward span start | ||
| t.recordEvent(TraceEvent{ | ||
| Name: "forward_to_peer", | ||
| Cat: "forward", | ||
| Ph: "B", | ||
| Tid: hashString(requestID) % 1000, | ||
| Args: map[string]interface{}{ | ||
| "request_id": requestID, | ||
| "target_node": targetNode, | ||
| "target_ip": targetIP, | ||
| "flow_id": flowID, | ||
| }, | ||
| }) | ||
|
|
||
| // Record flow event (arrow in Perfetto) | ||
| t.recordEvent(TraceEvent{ | ||
| Name: "forward_flow", | ||
| Cat: "forward", | ||
| Ph: "s", // Flow start | ||
| ID: flowID, | ||
| Tid: hashString(requestID) % 1000, | ||
| Scope: "forward", | ||
| }) | ||
|
|
||
| return flowID | ||
| } | ||
|
|
||
| // TraceForwardEnd records when forwarding completes | ||
| func (t *Tracer) TraceForwardEnd(requestID string, statusCode int, durationMs float64) { | ||
| if !t.IsActive() { | ||
| return | ||
| } | ||
|
|
||
| t.recordEvent(TraceEvent{ | ||
| Name: "forward_to_peer", | ||
| Cat: "forward", | ||
| Ph: "E", | ||
| Tid: hashString(requestID) % 1000, | ||
| Args: map[string]interface{}{ | ||
| "request_id": requestID, | ||
| "status_code": statusCode, | ||
| "duration_ms": durationMs, | ||
| }, | ||
| }) | ||
| } | ||
|
|
||
| // TraceLocalProcess records local SGLang processing | ||
| func (t *Tracer) TraceLocalProcessStart(requestID string) { | ||
| if !t.IsActive() { | ||
| return | ||
| } | ||
|
|
||
| t.recordEvent(TraceEvent{ | ||
| Name: "local_sglang", | ||
| Cat: "process", | ||
| Ph: "B", | ||
| Tid: hashString(requestID) % 1000, | ||
| Args: map[string]interface{}{ | ||
| "request_id": requestID, | ||
| }, | ||
| }) | ||
| } | ||
|
|
||
| func (t *Tracer) TraceLocalProcessEnd(requestID string, statusCode int) { | ||
| if !t.IsActive() { | ||
| return | ||
| } | ||
|
|
||
| t.recordEvent(TraceEvent{ | ||
| Name: "local_sglang", | ||
| Cat: "process", | ||
| Ph: "E", | ||
| Tid: hashString(requestID) % 1000, | ||
| Args: map[string]interface{}{ | ||
| "request_id": requestID, | ||
| "status_code": statusCode, | ||
| }, | ||
| }) | ||
| } | ||
|
|
||
| // TracePolicyDecision records a policy decision (instant event) | ||
| func (t *Tracer) TracePolicyDecision(requestID, policy, decision string, metadata map[string]interface{}) { | ||
| if !t.IsActive() { | ||
| return | ||
| } | ||
|
|
||
| args := map[string]interface{}{ | ||
| "request_id": requestID, | ||
| "policy": policy, | ||
| "decision": decision, | ||
| } | ||
| for k, v := range metadata { | ||
| args[k] = v | ||
| } | ||
|
|
||
| t.recordEvent(TraceEvent{ | ||
| Name: "policy_decision", | ||
| Cat: "policy", | ||
| Ph: "i", // Instant event | ||
| Tid: hashString(requestID) % 1000, | ||
| Args: args, | ||
| }) | ||
| } | ||
|
|
||
| // TraceCapacityCheck records capacity check result | ||
| func (t *Tracer) TraceCapacityCheck(requestID string, hasCapacity bool, runningReqs, waitingReqs int) { | ||
| if !t.IsActive() { | ||
| return | ||
| } | ||
|
|
||
| t.recordEvent(TraceEvent{ | ||
| Name: "capacity_check", | ||
| Cat: "lb", | ||
| Ph: "i", | ||
| Tid: hashString(requestID) % 1000, | ||
| Args: map[string]interface{}{ | ||
| "request_id": requestID, | ||
| "has_capacity": hasCapacity, | ||
| "running_reqs": runningReqs, | ||
| "waiting_reqs": waitingReqs, | ||
| }, | ||
| }) | ||
| } | ||
|
|
||
| // TraceTokenization records how long tokenization took (critical for GORGO overhead) | ||
| func (t *Tracer) TraceTokenization(requestID string, textLen, tokenCount int, durationUs int64) { | ||
| if !t.IsActive() { | ||
| return | ||
| } | ||
|
|
||
| t.recordEvent(TraceEvent{ | ||
| Name: "tokenize", | ||
| Cat: "gorgo", | ||
| Ph: "X", // Complete event with duration | ||
| Tid: hashString(requestID) % 1000, | ||
| Dur: durationUs, | ||
| Args: map[string]interface{}{ | ||
| "request_id": requestID, | ||
| "text_len": textLen, | ||
| "token_count": tokenCount, | ||
| "duration_us": durationUs, | ||
| }, | ||
| }) | ||
| } | ||
|
|
||
| // TraceGORGODecision records GORGO's routing decision with cost calculation | ||
| func (t *Tracer) TraceGORGODecision(requestID string, decision string, metadata map[string]interface{}) { | ||
| if !t.IsActive() { | ||
| return | ||
| } | ||
|
|
||
| args := map[string]interface{}{ | ||
| "request_id": requestID, | ||
| "decision": decision, | ||
| } | ||
| for k, v := range metadata { | ||
| args[k] = v | ||
| } | ||
|
|
||
| t.recordEvent(TraceEvent{ | ||
| Name: "gorgo_decision", | ||
| Cat: "gorgo", | ||
| Ph: "i", | ||
| Tid: hashString(requestID) % 1000, | ||
| Args: args, | ||
| }) | ||
| } | ||
|
|
||
| // ===================================================== | ||
| // DISTRIBUTED TRACING - Receive events from other nodes | ||
| // ===================================================== | ||
|
|
||
| // ReceiveRemoteEvents merges trace events from a forwarded request | ||
| func (t *Tracer) ReceiveRemoteEvents(events []TraceEvent) { | ||
| if !t.IsActive() { | ||
| return | ||
| } | ||
|
|
||
| t.sessionMu.RLock() | ||
| session := t.session | ||
| t.sessionMu.RUnlock() | ||
|
|
||
| if session == nil { | ||
| return | ||
| } | ||
|
|
||
| session.remoteEventsMu.Lock() | ||
| session.remoteEvents = append(session.remoteEvents, events...) | ||
| session.remoteEventsMu.Unlock() | ||
| } | ||
|
|
||
| // GetTraceHeaders returns headers to propagate trace context | ||
| func (t *Tracer) GetTraceHeaders(requestID string) map[string]string { | ||
| if !t.IsActive() { | ||
| return nil | ||
| } | ||
|
|
||
| t.sessionMu.RLock() | ||
| defer t.sessionMu.RUnlock() | ||
|
|
||
| if t.session == nil { | ||
| return nil | ||
| } | ||
|
|
||
| return map[string]string{ | ||
| "X-Trace-ID": t.session.ID, | ||
| "X-Trace-Request": requestID, | ||
| "X-Trace-Origin": t.nodeID, | ||
| } | ||
| } | ||
|
|
||
| // ===================================================== | ||
| // INTERNAL HELPERS | ||
| // ===================================================== | ||
|
|
||
| func (t *Tracer) recordEvent(event TraceEvent) { | ||
| t.sessionMu.RLock() | ||
| session := t.session | ||
| t.sessionMu.RUnlock() | ||
|
|
||
| if session == nil || !session.active.Load() { | ||
| return | ||
| } | ||
|
|
||
| // Set timestamp relative to trace start | ||
| event.Ts = time.Since(session.StartTime).Microseconds() | ||
| event.Pid = t.nodePID | ||
|
|
||
| session.eventsMu.Lock() | ||
| session.events = append(session.events, event) | ||
| session.eventsMu.Unlock() | ||
| } | ||
|
|
||
| // hashString creates a simple hash for consistent thread IDs | ||
| func hashString(s string) int { | ||
| h := 0 | ||
| for _, c := range s { | ||
| h = 31*h + int(c) | ||
| } | ||
| if h < 0 { | ||
| h = -h | ||
| } | ||
| return h | ||
| } |
There was a problem hiding this comment.
The pkg/trace/trace.go and pkg/cluster/trace.go files are identical duplicates (573 lines each). This violates the DRY principle and creates maintenance burden. One should be removed and the other should be imported where needed.
Comment on lines
+110
to
+111
| std::fs::set_permissions(socket_path, std::fs::Permissions::from_mode(0o777)) | ||
| .expect("Failed to set socket permissions"); |
There was a problem hiding this comment.
In the Rust tokenizer sidecar, the socket permissions are set to 0o777 (world-readable/writable). This is a security risk as any user on the system can connect to the tokenizer socket and potentially cause denial of service or other issues. Consider using more restrictive permissions like 0o660 and ensuring proper group ownership.
benchmark/geo_proxy.py
Outdated
| ) | ||
|
|
||
| except Exception as e: | ||
| status = "error" |
There was a problem hiding this comment.
Variable status is not used.
benchmark/geo_proxy.py
Outdated
Comment on lines
236
to
238
| self.logger.error(f"[{request_id}] Request failed: {e}") | ||
| return web.json_response( | ||
| {"error": {"message": str(e), "type": "server_error"}}, |
There was a problem hiding this comment.
Variable error_msg is not used.
Suggested change
| self.logger.error(f"[{request_id}] Request failed: {e}") | |
| return web.json_response( | |
| {"error": {"message": str(e), "type": "server_error"}}, | |
| self.logger.error(f"[{request_id}] Request failed: {error_msg}") | |
| return web.json_response( | |
| {"error": {"message": error_msg, "type": "server_error"}}, |
| import sys | ||
| import argparse | ||
| from collections import defaultdict | ||
| from typing import Dict, List, Tuple, Optional |
There was a problem hiding this comment.
Import of 'Optional' is not used.
Suggested change
| from typing import Dict, List, Tuple, Optional | |
| from typing import Dict, List, Tuple |
| import time | ||
| import uuid | ||
| from pathlib import Path | ||
| from typing import Any, Dict, List, Optional, Tuple |
There was a problem hiding this comment.
Import of 'Optional' is not used.
Import of 'List' is not used.
Suggested change
| from typing import Any, Dict, List, Optional, Tuple | |
| from typing import Any, Dict, Tuple |
Comment on lines
+134
to
+151
| # Build GuideLLM command | ||
| GUIDELLM_CMD="guidellm benchmark \ | ||
| --target http://localhost:$PROXY_PORT/v1 \ | ||
| --profile $PROFILE \ | ||
| --max-seconds $DURATION \ | ||
| --data $SCRIPT_DIR/wildchat_guidellm.jsonl \ | ||
| --data-column-mapper '{\"text_column\": \"prompt\"}' \ | ||
| --output-path $OUTPUT_DIR" | ||
|
|
||
| # Add rate for profiles that need it | ||
| if [ "$PROFILE" = "poisson" ] || [ "$PROFILE" = "constant" ] || [ "$PROFILE" = "concurrent" ] || [ "$PROFILE" = "throughput" ]; then | ||
| GUIDELLM_CMD="$GUIDELLM_CMD --rate $RATE" | ||
| fi | ||
|
|
||
| echo "Running: $GUIDELLM_CMD" | ||
| echo "" | ||
|
|
||
| eval $GUIDELLM_CMD |
There was a problem hiding this comment.
The use of eval $GUIDELLM_CMD with command-line-controlled variables like PROFILE, RATE, DURATION, and OUTPUT_DIR creates a command injection risk: a crafted argument (for example, --profile 'poisson; rm -rf ~') would cause arbitrary shell commands to execute in the context of whoever runs this script. To prevent this, avoid eval entirely and invoke guidellm directly using a properly quoted argument list (e.g., building an array and calling it without eval) so user-supplied values cannot break out of their argument positions.
Suggested change
| # Build GuideLLM command | |
| GUIDELLM_CMD="guidellm benchmark \ | |
| --target http://localhost:$PROXY_PORT/v1 \ | |
| --profile $PROFILE \ | |
| --max-seconds $DURATION \ | |
| --data $SCRIPT_DIR/wildchat_guidellm.jsonl \ | |
| --data-column-mapper '{\"text_column\": \"prompt\"}' \ | |
| --output-path $OUTPUT_DIR" | |
| # Add rate for profiles that need it | |
| if [ "$PROFILE" = "poisson" ] || [ "$PROFILE" = "constant" ] || [ "$PROFILE" = "concurrent" ] || [ "$PROFILE" = "throughput" ]; then | |
| GUIDELLM_CMD="$GUIDELLM_CMD --rate $RATE" | |
| fi | |
| echo "Running: $GUIDELLM_CMD" | |
| echo "" | |
| eval $GUIDELLM_CMD | |
| # Build GuideLLM command as an argument array to avoid command injection | |
| GUIDELLM_ARGS=( | |
| benchmark | |
| --target "http://localhost:$PROXY_PORT/v1" | |
| --profile "$PROFILE" | |
| --max-seconds "$DURATION" | |
| --data "$SCRIPT_DIR/wildchat_guidellm.jsonl" | |
| --data-column-mapper '{"text_column": "prompt"}' | |
| --output-path "$OUTPUT_DIR" | |
| ) | |
| # Add rate for profiles that need it | |
| if [ "$PROFILE" = "poisson" ] || [ "$PROFILE" = "constant" ] || [ "$PROFILE" = "concurrent" ] || [ "$PROFILE" = "throughput" ]; then | |
| GUIDELLM_ARGS+=( --rate "$RATE" ) | |
| fi | |
| # Show the command being run | |
| printf 'Running: guidellm' | |
| for arg in "${GUIDELLM_ARGS[@]}"; do | |
| printf ' %q' "$arg" | |
| done | |
| printf '\n\n' | |
| guidellm "${GUIDELLM_ARGS[@]}" |
Owner
Author
|
Sounds good, thanks copilot.
1. Get rid of the entire benchmark folder for this PR
2. Productionize the tokenizer policies. Add graceful error handling,
restrict socket access from local processes only, and fix the race
condition from two concurrent requests connecting.
3. get rid of remove instance from config
…On Wed, Jan 28, 2026 at 9:24 PM Copilot ***@***.***> wrote:
***@***.**** commented on this pull request.
Pull request overview
This PR introduces a comprehensive load balancing infrastructure for
distributed GPU cluster management. It adds event-driven tracing
(Perfetto-compatible), tokenizer sidecar integration (Rust), SGLang cluster
management, and multiple load balancing policies (GORGO, prefix-tree,
least-loaded). The PR also refactors instance state management to use
Lambda API as the source of truth instead of local database storage.
*Changes:*
- Added distributed tracing system with Perfetto format support
- Implemented Rust tokenizer sidecar with Unix socket communication
- Created cluster management infrastructure for SGLang deployments
- Added load balancing policies and configuration system
- Removed local DB dependency for instance state, using Lambda API
instead
- Added comprehensive benchmarking tools using WildChat-1M dataset
Reviewed changes
Copilot reviewed 39 out of 42 changed files in this pull request and
generated 12 comments.
Show a summary per file
File Description
pkg/trace/trace.go New event-driven tracing system (Perfetto format)
pkg/cluster/trace.go Duplicate of pkg/trace/trace.go
pkg/tokenizer/*.go Tokenizer sidecar management and client
pkg/tokenizer/main.rs Rust tokenizer service with Unix socket
pkg/sglang/sglang.go SGLang Docker deployment and diagnostics
pkg/cluster/*.go Cluster management, metrics, build/deploy
pkg/config/constants.go Centralized configuration constants
pkg/remote/client.go Refactored SSH key lookup from Lambda API
cmd/*.go New commands for cluster, tokenizer, LB management
benchmark/*.py Geo-distributed benchmarking tools
go.mod Added tokenizers dependency (with conflicts)
------------------------------
💡 Add Copilot custom instructions
<http:///atoniolo76/gotoni/new/main/.github/instructions?filename=*.instructions.md>
for smarter, more guided reviews. Learn how to get started
<https://docs.github.com/en/copilot/customizing-copilot/adding-repository-custom-instructions-for-github-copilot>
.
------------------------------
In pkg/tokenizer/main.rs
<#4 (comment)>:
> + .expect("Failed to download tokenizer.json");
+
+ if !response.status().is_success() {
+ panic!("Failed to download tokenizer: HTTP {}", response.status());
+ }
+
+ let bytes = response.bytes().await
+ .expect("Failed to read tokenizer bytes");
+
+ // Cache the tokenizer
+ std::fs::write(TOKENIZER_CACHE_PATH, &bytes)
+ .expect("Failed to cache tokenizer");
+
+ println!("Tokenizer cached to: {}", TOKENIZER_CACHE_PATH);
+
+ Tokenizer::from_bytes(&bytes)
+ .expect("Failed to parse tokenizer.json")
+}
+
+#[tokio::main]
+async fn main() {
+ println!("Loading tokenizer model: {}", MODEL_ID);
+
+ // Load the Mistral tokenizer
+ let tokenizer: Arc<Tokenizer> = Arc::new(load_or_download_tokenizer().await);
+ println!("Tokenizer loaded successfully");
+
+ // Build the router with shared state
+ let tk_for_count = Arc::clone(&tokenizer);
+
+ let app = Router::new()
+ .route("/count", post(move |Json(payload): Json<TokenizeRequest>| {
+ let tk: Arc<Tokenizer> = Arc::clone(&tk_for_count);
+ async move {
+ let encoding = tk.encode(payload.text, false).expect("Tokenization failed");
+ Json(TokenizeResponse {
+ count: encoding.get_ids().len(),
+ })
+ }
+ }))
+ .route("/health", get(|| async {
+ Json(HealthResponse {
+ status: "ok".to_string(),
+ model: MODEL_ID.to_string(),
+ })
+ }));
+
+ // Unix Domain Socket path
+ let socket_path = "/tmp/tokenizer.sock";
+
+ // Clean up old socket if it exists
+ let _ = std::fs::remove_file(socket_path);
+
+ println!("Starting tokenizer sidecar on {}", socket_path);
+
+ let listener = UnixListener::bind(socket_path).expect("Failed to bind to socket");
+
+ // Make socket world-readable/writable
+ std::fs::set_permissions(socket_path, std::fs::Permissions::from_mode(0o777))
+ .expect("Failed to set socket permissions");
+
+ // Serve using hyper with Unix socket
+ loop {
+ let (stream, _) = listener.accept().await.expect("Accept failed");
The tokenizer sidecar uses expect() for error handling in several places
(lines 52, 59, 63, 68, 86, 107, 111, 115), which will panic the entire
process if the operation fails. For a production service, these should use
proper error handling with ? operator or unwrap_or_else() with logging to
allow graceful degradation rather than crashing.
------------------------------
In pkg/remote/client.go
<#4 (comment)>:
> +// RemoveInstanceFromConfig removes an instance from the config
+// Deprecated: Instance state is now managed by Lambda API, not local DB.
+// This function is kept for backwards compatibility but does nothing meaningful.
+func RemoveInstanceFromConfig(instanceID string) error {
+ // Instance state is managed by Lambda API, not local DB
+ // No action needed - when an instance is terminated via Lambda API,
+ // it will no longer appear in ListRunningInstances
+ return nil
}
The RemoveInstanceFromConfig function is deprecated but returns no error
and does nothing. Callers expecting this function to clean up instance
state will silently fail. Consider either removing this function entirely
or returning an error indicating it's no longer supported to make the
deprecation explicit.
------------------------------
In pkg/cluster/tokenizer.go
<#4 (comment)>:
> + tokenizerMu.RLock()
+ available := tokenizerAvailable
+ tokenizerMu.RUnlock()
+
+ if !available {
+ // Try to reconnect periodically
+ go func() {
+ tokenizerMu.Lock()
+ defer tokenizerMu.Unlock()
+ if !tokenizerAvailable {
+ tokenizerAvailable = checkTokenizerHealth()
+ }
+ }()
+ return fallbackTokenCount(text)
The pkg/cluster/tokenizer.go file has a potential race condition. The
tokenizerAvailable flag is read without lock on line 94 and checked again
on line 102-103, but another goroutine could modify it between these
checks. The reconnection goroutine spawned on line 99 should be
synchronized to prevent multiple concurrent reconnection attempts.
------------------------------
In pkg/cluster/trace.go
<#4 (comment)>:
> +
+// TraceGORGODecision records GORGO's routing decision with cost calculation
+func (t *Tracer) TraceGORGODecision(requestID string, decision string, metadata map[string]interface{}) {
+ if !t.IsActive() {
+ return
+ }
+
+ args := map[string]interface{}{
+ "request_id": requestID,
+ "decision": decision,
+ }
+ for k, v := range metadata {
+ args[k] = v
+ }
+
+ t.recordEvent(TraceEvent{
+ Name: "gorgo_decision",
+ Cat: "gorgo",
+ Ph: "i",
+ Tid: hashString(requestID) % 1000,
+ Args: args,
+ })
+}
+
+// =====================================================
+// DISTRIBUTED TRACING - Receive events from other nodes
+// =====================================================
+
+// ReceiveRemoteEvents merges trace events from a forwarded request
+func (t *Tracer) ReceiveRemoteEvents(events []TraceEvent) {
+ if !t.IsActive() {
+ return
+ }
+
+ t.sessionMu.RLock()
+ session := t.session
+ t.sessionMu.RUnlock()
+
+ if session == nil {
+ return
+ }
+
+ session.remoteEventsMu.Lock()
+ session.remoteEvents = append(session.remoteEvents, events...)
+ session.remoteEventsMu.Unlock()
+}
+
+// GetTraceHeaders returns headers to propagate trace context
+func (t *Tracer) GetTraceHeaders(requestID string) map[string]string {
+ if !t.IsActive() {
+ return nil
+ }
+
+ t.sessionMu.RLock()
+ defer t.sessionMu.RUnlock()
+
+ if t.session == nil {
+ return nil
+ }
+
+ return map[string]string{
+ "X-Trace-ID": t.session.ID,
+ "X-Trace-Request": requestID,
+ "X-Trace-Origin": t.nodeID,
+ }
+}
+
+// =====================================================
+// INTERNAL HELPERS
+// =====================================================
+
+func (t *Tracer) recordEvent(event TraceEvent) {
+ t.sessionMu.RLock()
+ session := t.session
+ t.sessionMu.RUnlock()
+
+ if session == nil || !session.active.Load() {
+ return
+ }
+
+ // Set timestamp relative to trace start
+ event.Ts = time.Since(session.StartTime).Microseconds()
+ event.Pid = t.nodePID
+
+ session.eventsMu.Lock()
+ session.events = append(session.events, event)
+ session.eventsMu.Unlock()
+}
+
+// hashString creates a simple hash for consistent thread IDs
+func hashString(s string) int {
+ h := 0
+ for _, c := range s {
+ h = 31*h + int(c)
+ }
+ if h < 0 {
+ h = -h
+ }
+ return h
+}
The pkg/trace/trace.go and pkg/cluster/trace.go files are identical
duplicates (573 lines each). This violates the DRY principle and creates
maintenance burden. One should be removed and the other should be imported
where needed.
------------------------------
In pkg/tokenizer/main.rs
<#4 (comment)>:
> + std::fs::set_permissions(socket_path, std::fs::Permissions::from_mode(0o777))
+ .expect("Failed to set socket permissions");
In the Rust tokenizer sidecar, the socket permissions are set to 0o777
(world-readable/writable). This is a security risk as any user on the
system can connect to the tokenizer socket and potentially cause denial of
service or other issues. Consider using more restrictive permissions like
0o660 and ensuring proper group ownership.
------------------------------
In benchmark/geo_proxy.py
<#4 (comment)>:
> + )
+ await response.prepare(request)
+
+ # Stream response chunks
+ first_chunk = True
+ async for chunk in resp.content.iter_any():
+ if first_chunk:
+ ttft = time.time() - start_time
+ first_chunk = False
+ await response.write(chunk)
+ total_tokens += 1 # Approximate
+
+ await response.write_eof()
+
+ except asyncio.TimeoutError:
+ status = "timeout"
Variable status is not used.
------------------------------
In benchmark/geo_proxy.py
<#4 (comment)>:
> + await response.write(chunk)
+ total_tokens += 1 # Approximate
+
+ await response.write_eof()
+
+ except asyncio.TimeoutError:
+ status = "timeout"
+ error_msg = f"Request timed out after {REQUEST_TIMEOUT}s"
+ self.logger.error(f"[{request_id}] {error_msg}")
+ return web.json_response(
+ {"error": {"message": error_msg, "type": "timeout_error"}},
+ status=504
+ )
+
+ except Exception as e:
+ status = "error"
Variable status is not used.
------------------------------
In benchmark/geo_proxy.py
<#4 (comment)>:
> + self.logger.error(f"[{request_id}] Request failed: {e}")
+ return web.json_response(
+ {"error": {"message": str(e), "type": "server_error"}},
Variable error_msg is not used.
⬇️ Suggested change
- self.logger.error(f"[{request_id}] Request failed: {e}")
- return web.json_response(
- {"error": {"message": str(e), "type": "server_error"}},
+ self.logger.error(f"[{request_id}] Request failed: {error_msg}")
+ return web.json_response(
+ {"error": {"message": error_msg, "type": "server_error"}},
------------------------------
In benchmark/analyze_trace.py
<#4 (comment)>:
> +Usage:
+ # Fetch and analyze trace from a running LB
+ curl http://192.9.140.201:8000/lb/trace/stop | python analyze_trace.py
+
+ # Analyze a saved trace file
+ python analyze_trace.py trace.json
+
+ # Or use the built-in fetch
+ python analyze_trace.py --fetch http://192.9.140.201:8000
+"""
+
+import json
+import sys
+import argparse
+from collections import defaultdict
+from typing import Dict, List, Tuple, Optional
Import of 'Optional' is not used.
⬇️ Suggested change
-from typing import Dict, List, Tuple, Optional
+from typing import Dict, List, Tuple
------------------------------
In benchmark/geo_proxy.py
<#4 (comment)>:
> +Usage:
+ python geo_proxy.py [--port 9000] [--lookup-file wildchat_location_lookup.json]
+
+Then point GuideLLM at:
+ guidellm benchmark --target http://localhost:9000/v1 ...
+"""
+import argparse
+import asyncio
+import hashlib
+import json
+import logging
+import math
+import time
+import uuid
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Tuple
Import of 'Optional' is not used.
Import of 'List' is not used.
⬇️ Suggested change
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Any, Dict, Tuple
------------------------------
In benchmark/run_guidellm_geo.sh
<#4 (comment)>:
> +# Build GuideLLM command
+GUIDELLM_CMD="guidellm benchmark \
+ --target http://localhost:$PROXY_PORT/v1 \
+ --profile $PROFILE \
+ --max-seconds $DURATION \
+ --data $SCRIPT_DIR/wildchat_guidellm.jsonl \
+ --data-column-mapper '{\"text_column\": \"prompt\"}' \
+ --output-path $OUTPUT_DIR"
+
+# Add rate for profiles that need it
+if [ "$PROFILE" = "poisson" ] || [ "$PROFILE" = "constant" ] || [ "$PROFILE" = "concurrent" ] || [ "$PROFILE" = "throughput" ]; then
+ GUIDELLM_CMD="$GUIDELLM_CMD --rate $RATE"
+fi
+
+echo "Running: $GUIDELLM_CMD"
+echo ""
+
+eval $GUIDELLM_CMD
The use of eval $GUIDELLM_CMD with command-line-controlled variables like
PROFILE, RATE, DURATION, and OUTPUT_DIR creates a command injection risk:
a crafted argument (for example, --profile 'poisson; rm -rf ~') would
cause arbitrary shell commands to execute in the context of whoever runs
this script. To prevent this, avoid eval entirely and invoke guidellm
directly using a properly quoted argument list (e.g., building an array and
calling it without eval) so user-supplied values cannot break out of
their argument positions.
⬇️ Suggested change
-# Build GuideLLM command
-GUIDELLM_CMD="guidellm benchmark \
- --target http://localhost:$PROXY_PORT/v1 \
- --profile $PROFILE \
- --max-seconds $DURATION \
- --data $SCRIPT_DIR/wildchat_guidellm.jsonl \
- --data-column-mapper '{\"text_column\": \"prompt\"}' \
- --output-path $OUTPUT_DIR"
-
-# Add rate for profiles that need it
-if [ "$PROFILE" = "poisson" ] || [ "$PROFILE" = "constant" ] || [ "$PROFILE" = "concurrent" ] || [ "$PROFILE" = "throughput" ]; then
- GUIDELLM_CMD="$GUIDELLM_CMD --rate $RATE"
-fi
-
-echo "Running: $GUIDELLM_CMD"
-echo ""
-
-eval $GUIDELLM_CMD
+# Build GuideLLM command as an argument array to avoid command injection
+GUIDELLM_ARGS=(
+ benchmark
+ --target "http://localhost:$PROXY_PORT/v1"
+ --profile "$PROFILE"
+ --max-seconds "$DURATION"
+ --data "$SCRIPT_DIR/wildchat_guidellm.jsonl"
+ --data-column-mapper '{"text_column": "prompt"}'
+ --output-path "$OUTPUT_DIR"
+)
+
+# Add rate for profiles that need it
+if [ "$PROFILE" = "poisson" ] || [ "$PROFILE" = "constant" ] || [ "$PROFILE" = "concurrent" ] || [ "$PROFILE" = "throughput" ]; then
+ GUIDELLM_ARGS+=( --rate "$RATE" )
+fi
+
+# Show the command being run
+printf 'Running: guidellm'
+for arg in "${GUIDELLM_ARGS[@]}"; do
+ printf ' %q' "$arg"
+done
+printf '\n\n'
+
+guidellm "${GUIDELLM_ARGS[@]}"
—
Reply to this email directly, view it on GitHub
<#4 (review)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AKPYOD5UGWICIQS2MALIVED4JFVNJAVCNFSM6AAAAACTIBXQFSVHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMZTOMRQGE3TMMBYG4>
.
You are receiving this because you authored the thread.Message ID:
***@***.***>
|
…o76/gotoni into benchmark-load-balancing
Adds a new CLI command to visualize network latencies between cluster nodes: - `gotoni lb latency --all` shows latencies from all nodes to their peers - `--matrix` flag displays as a latency matrix for easier comparison - Helps understand inter-region transport costs for GORGO/GORGO2 decisions 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
SSH keys are now looked up directly from ~/.ssh directory by name, without requiring the local database. The Lambda API provides the SSH key name for each instance, and we find the corresponding .pem file in the user's ssh directory. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
- Add pkg/proxy/proxy.go: Centralized HTTP proxy with GORGO-style routing - Global request tracking across all SGLang servers - Cost calculation: queued_tokens * ms_per_token + running_tokens * ms_per_token * 0.5 - EWMA latency tracking with dedicated probing - Prefix tree for KV cache-aware routing - Add cmd/proxy.go: CLI commands for proxy management - proxy start/stop/status: Local proxy control - proxy deploy: Build, upload, and start proxy on remote instance - proxy servers add/remove: Manage SGLang server pool - proxy tune: Adjust GORGO parameters at runtime - proxy latency: Display latencies to all servers - --remote flag: Resolve instance names via Lambda API 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
…+ lb installation, Update README for Research
Updated README to include './' before commands for consistency.
…equests, updated constants and README
…equests, and updated README
…o76/gotoni into benchmark-load-balancing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Create infrastructure for clusters, load balancing, tokenizers and more. Overhaul the existing db system with stateless calls from lambda's backend. Add a bunch of commands for interacting with various components specific to GORGO load balancing. Add policies such as prefix-tree and least-load for inference users. Integrate @Orgo as a provider.