dist/net_rdma: batch transfer of entries if possible

dist/net_rdma.c

@@ -42,6 +42,7 @@
 
 static const uint64_t kPollReportThreshold = 128;
 static const uint64_t kCleanReportThreshold = 128;
+static const uint64_t kPollMax = 8;
 
 const char *shm_path = NULL;
 size_t shm_size = 256 * 1024 * 1024ULL;  // 256MB
@@ -343,7 +344,9 @@ static int PeerEvent(struct Peer *peer, uint32_t events) {
 static inline void PollPeerTransfer(struct Peer *peer, bool *report) {
   // XXX: consider batching this to forward multiple entries at once if possible
 
-  void *entry = (peer->local_base + peer->local_pos * peer->local_elen);
+  size_t n;
+  for (n = 0; n < kPollMax && peer->local_pos + n < peer->local_enum; n++) {
+    void *entry = (peer->local_base + (peer->local_pos + n) * peer->local_elen);
     bool ready;
     if (peer->is_dev) {
       struct SimbricksProtoNetD2NDummy *d2n = entry;
@@ -354,10 +357,13 @@ static inline void PollPeerTransfer(struct Peer *peer, bool *report) {
       ready = (n2d->own_type & SIMBRICKS_PROTO_NET_N2D_OWN_MASK) ==
           SIMBRICKS_PROTO_NET_N2D_OWN_DEV;
     }
+    if (!ready)
+      break;
+  }
 
-  if (ready) {
-    RdmaPassEntry(peer);
-    peer->local_pos += 1;
+  if (n > 0) {
+    RdmaPassEntry(peer, n);
+    peer->local_pos += n;
     if (peer->local_pos >= peer->local_enum)
       peer->local_pos -= peer->local_enum;
 

dist/net_rdma.h

@@ -106,7 +106,7 @@ int PeerReport(struct Peer *peer, uint32_t written_pos, uint32_t clean_pos);
 int RdmaListen(struct sockaddr_in *addr);
 int RdmaConnect(struct sockaddr_in *addr);
 int RdmaPassIntro(struct Peer *peer);
-int RdmaPassEntry(struct Peer *peer);
+int RdmaPassEntry(struct Peer *peer, uint32_t n);
 int RdmaPassReport();
 int RdmaEvent();
 

dist/rdma.c

@@ -550,7 +550,7 @@ int RdmaPassIntro(struct Peer *peer) {
   return 0;
 }
 
-int RdmaPassEntry(struct Peer *peer) {
+int RdmaPassEntry(struct Peer *peer, uint32_t n) {
 #ifdef RDMA_DEBUG
   fprintf(stderr, "RdmaPassEntry(%s,%u)\n", peer->sock_path, peer->local_pos);
   fprintf(stderr, "  remote_base=%lx local_base=%p\n", peer->remote_base,
@@ -560,7 +560,7 @@ int RdmaPassEntry(struct Peer *peer) {
   uint64_t pos = peer->local_pos * peer->local_elen;
   struct ibv_sge sge;
   sge.addr = (uintptr_t) (peer->local_base + pos);
-  sge.length = peer->local_elen;
+  sge.length = peer->local_elen * n;
   sge.lkey = peer->shm_mr->lkey;
 
   struct ibv_send_wr send_wr = { };