Fix simulation hang when QP allocation in SendFlow produces 0-byte or mismatched-size QPs

[Copilot]

When _QPS_PER_CONNECTION_ exceeds the available data or when custom per-QP allocation is used, two bugs in entry.h combine to permanently stall the simulation: counters in waiting_to_sent_callback never reach zero, and the upper-layer send-complete callback is never fired.

Root causes & fixes

SendFlow: 0-byte QPs inflate the actual sent total

Previously, when maxPacketCount < _QPS_PER_CONNECTION_, surplus loop iterations computed real_PacketCount = 0. Port allocation, sender_src_port_map insertion, and both waiting-counter increments had already happened before the check — which then bumped the size to 1 rather than skipping. This made all_sent_chunksize > maxPacketCount, breaking the equality check below.

Fix: Move the zero check before any resource allocation and use continue:

// Before: silently inflated total bytes sent
if(real_PacketCount == 0) real_PacketCount = 1;

// After: skip the QP entirely — no port, no map entry, no counter bump
if(real_PacketCount == 0) {
    NcclLog->writeLog(NcclLogLevel::DEBUG, " [SendFlow] skipping 0-byte QP ...");
    continue;
}
uint32_t port = portNumber[src][dst]++;

notify_sender_sending_finished: strict equality blocks the callback

The condition t2.count == message_size meant any size mismatch (from the above bug, or from integer rounding in custom allocation) silently dropped the callback, leaving the simulation deadlocked. The receiver-side counterpart (notify_receiver_receive_data) already handles message_size > t2.count correctly.

Fix: Relax to >=, consistent with receiver-side logic:

// Before
if (t2.count == message_size) { ... }

// After
if (message_size >= t2.count) { ... }

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[Copilot]

Pull request overview

Fixes a simulation deadlock in the ns-3 network frontend when per-connection QP splitting can produce 0-byte QPs or when the total sent size doesn’t exactly match the requested message size, preventing send-complete callbacks from firing.

Changes:

• Skip creation/allocation of QPs that would carry 0 bytes in SendFlow.
• Relax sender completion condition in notify_sender_sending_finished from strict equality to >= to ensure callbacks fire on size mismatches.

Comments suppressed due to low confidence (1)

astra-sim-alibabacloud/astra-sim/network_frontend/ns3/entry.h:269

• The mismatch path here is now specifically message_size < t2.count, but the error log still says "msg size !=" and it doesn’t include the expected t2.count (or any flow/tag identifier). Also, the format string ends with flow_id but no value is actually printed. Updating the log to report both actual/expected counts plus flowTag.current_flow_id/tag will make stalls/mismatches diagnosable.

      if (message_size >= t2.count) {
        sentHash.erase(make_pair(tag, make_pair(sender_node, receiver_node)));
        if (nodeHash.find(make_pair(sender_node, 0)) == nodeHash.end()) {
          nodeHash[make_pair(sender_node, 0)] = message_size;
        } else {
          nodeHash[make_pair(sender_node, 0)] += message_size;
        }
        #ifdef NS3_MTP
        cs.ExitSection();
        #endif
        t2.msg_handler(t2.fun_arg);
        goto sender_end_1st_section;
      }else{
        NcclLog->writeLog(NcclLogLevel::ERROR,"sentHash msg size != sender_node %d receiver_node %d message_size %lu flow_id ",sender_node,receiver_node,message_size);
      }

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.

[Copilot]

Skipping real_PacketCount == 0 QPs prevents the counter/map inflation, but it also means a 0-byte send (maxPacketCount==0) will schedule no QPs and therefore never call send_finish() / notify_sender_sending_finished(). Since AstraSimNetwork::sim_send() always inserts an entry into sentHash, a 0-byte send can now deadlock waiting for a completion that will never happen (e.g., Ring sets msg_size = data_size / nodes_in_ring, which can be 0 when data_size < nodes_in_ring). Consider handling maxPacketCount==0 (or “no QPs scheduled” after the loop) by immediately completing the send: erase the sentHash entry for (tag,src,dst) and invoke the upper-layer handler (or call notify_sender_sending_finished(src,dst,0,request->flowTag)).