add NIC_TRAFFIC_CLASS and NIC_SERVICE_LEVEL env vars for DSCP marking (#315)

Adds support for marking RoCE/IB traffic with specific DSCP/QoS values.

- NIC_TRAFFIC_CLASS (default=0): sets the DSCP/traffic class byte in the
  RoCE GRH (grh.traffic_class) when transitioning QPs to RTR state.
- NIC_SERVICE_LEVEL (default=0): sets the IB service level (ah_attr.sl)
  on QPs. This applies to IB and RoCE connections.
- NicOptions: I added uint8_t serviceLevel and uint8_t trafficClass fields
- TransitionQpToRtr(): accepts trafficClass and serviceLevel as parameters;
  sets grh.traffic_class (RoCE only) and ah_attr.sl (all QP types)

---------

Co-authored-by: Pak Nin Lui <paklui@smc300x-ccs-aus-gpuf2c9.prov.aus.ccs.cpe.ice.amd.com>

Author: paklui

CHANGELOG.md

@@ -5,6 +5,8 @@ Documentation for TransferBench is available at
 
 ## v1.67.00
 ### Added
+- Added NIC_TRAFFIC_CLASS to set the DSCP/traffic class byte in the RoCE GRH for QPs (RoCE only)
+- Added NIC_SERVICE_LEVEL to set the IB service level (sl) for QPs (IB and RoCE)
 - Initial support for pod communication.  Requires compatible hardware / ROCm version and subject to further testing
   - This potentially enables GFX/DMA executors to access SRC/DST memory locations on GPUs within the same pod
   - Pod membership requires amd-smi however can be skipped by setting TB_FORCE_SINGLE_POD=1

src/client/EnvVars.hpp

@@ -120,6 +120,8 @@ class EnvVars
   int nicChunkBytes;                 // Number of bytes to send per chunk for RDMA operations
   int nicCqPollBatch;                // Number of CQ entries to poll per ibv_poll_cq call
   int nicRelaxedOrder;               // Use relaxed ordering for RDMA
+  int nicServiceLevel;               // IB service level (sl) for InfiniBand QPs
+  int nicTrafficClass;               // DSCP/traffic class byte for RoCE GRH
   int roceVersion;                   // RoCE version number
 
   // Developer features
@@ -185,6 +187,18 @@ class EnvVars
     nicChunkBytes     = GetEnvVar("NIC_CHUNK_BYTES"     , 1073741824);
     nicCqPollBatch    = GetEnvVar("NIC_CQ_POLL_BATCH"   , 4);
     nicRelaxedOrder   = GetEnvVar("NIC_RELAX_ORDER"     , 1);
+    nicServiceLevel   = GetEnvVar("NIC_SERVICE_LEVEL"   , 0);
+    nicTrafficClass   = GetEnvVar("NIC_TRAFFIC_CLASS"   , 0);
+
+    // Check that NIC service level and traffic class are in valid ranges
+    if (nicServiceLevel < 0 || nicServiceLevel > 15) {
+      printf("[ERROR] NIC_SERVICE_LEVEL must be in range 0..15 (got %d)\n", nicServiceLevel);
+      exit(1);
+    }
+    if (nicTrafficClass < 0 || nicTrafficClass > 255) {
+      printf("[ERROR] NIC_TRAFFIC_CLASS must be in range 0..255 (got %d)\n", nicTrafficClass);
+      exit(1);
+    }
 
     gpuMaxHwQueues    = GetEnvVar("GPU_MAX_HW_QUEUES"   , 4);
 
@@ -366,6 +380,8 @@ class EnvVars
     printf(" NIC_CHUNK_BYTES     - Number of bytes to send at a time using NIC (default = 1GB)\n");
     printf(" NIC_CQ_POLL_BATCH   - Number of CQ entries to poll per ibv_poll_cq call (default = 4)\n");
     printf(" NIC_RELAX_ORDER     - Set to non-zero to use relaxed ordering\n");
+    printf(" NIC_SERVICE_LEVEL   - IB service level (sl) for InfiniBand QPs (default=0)\n");
+    printf(" NIC_TRAFFIC_CLASS   - DSCP/traffic class byte for RoCE GRH (default=0)\n");
 #endif
     printf(" NUM_ITERATIONS      - # of timed iterations per test. If negative, run for this many seconds instead\n");
     printf(" NUM_SUBITERATIONS   - # of sub-iterations to run per iteration. Must be non-negative\n");
@@ -504,6 +520,10 @@ class EnvVars
           "Polling %d CQ entries per ibv_poll_cq call", nicCqPollBatch);
     Print("NIC_RELAX_ORDER", nicRelaxedOrder,
           "Using %s ordering for NIC RDMA", nicRelaxedOrder ? "relaxed" : "strict");
+    Print("NIC_SERVICE_LEVEL", nicServiceLevel,
+          "IB service level (sl) set to %d", nicServiceLevel);
+    Print("NIC_TRAFFIC_CLASS", nicTrafficClass,
+          "RoCE traffic class (DSCP) set to %d", nicTrafficClass);
 #endif
     Print("NUM_ITERATIONS", numIterations,
           (numIterations == 0) ? "Running infinitely" :
@@ -725,6 +745,8 @@ class EnvVars
     cfg.nic.ibPort                 = ibPort;
     cfg.nic.ipAddressFamily        = ipAddressFamily;
     cfg.nic.useRelaxedOrder        = nicRelaxedOrder;
+    cfg.nic.serviceLevel           = nicServiceLevel;
+    cfg.nic.trafficClass           = nicTrafficClass;
     cfg.nic.roceVersion            = roceVersion;
 
     return cfg;

src/header/TransferBench.hpp

@@ -270,6 +270,8 @@ namespace TransferBench
     int         queueSize       = 100;          ///< Completion queue size
     int         roceVersion     = 2;            ///< RoCE version (used for auto GID detection)
     int         useRelaxedOrder = 1;            ///< Use relaxed ordering
+    uint8_t     serviceLevel    = 0;            ///< IB service level (sl) for InfiniBand QPs
+    uint8_t     trafficClass    = 0;            ///< DSCP/traffic class byte for RoCE GRH
     int         useNuma         = 0;            ///< Switch to closest numa thread for execution
   };
 
@@ -2000,6 +2002,8 @@ namespace {
       if (nic.maxSendWorkReq  != cfg.nic.maxSendWorkReq)  ADD_ERROR("cfg.nic.maxSendWorkReq");
       // nic.queueSize   is permitted to be different across ranks
       if (nic.roceVersion     != cfg.nic.roceVersion)     ADD_ERROR("cfg.nic.roceVersion");
+      if (nic.serviceLevel    != cfg.nic.serviceLevel)    ADD_ERROR("cfg.nic.serviceLevel");
+      if (nic.trafficClass    != cfg.nic.trafficClass)    ADD_ERROR("cfg.nic.trafficClass");
       if (nic.useRelaxedOrder != cfg.nic.useRelaxedOrder) ADD_ERROR("cfg.nic.useRelaxedOrder");
       if (nic.useNuma         != cfg.nic.useNuma)         ADD_ERROR("cfg.nic.useNuma");
     }
@@ -3284,7 +3288,9 @@ static bool IsConfiguredGid(union ibv_gid const& gid)
                                      ConnInfo const& connInfo,
                                      uint8_t  const& port,
                                      bool     const& isRoCE,
-                                     ibv_mtu  const& mtu)
+                                     ibv_mtu  const& mtu,
+                                     uint8_t  const& trafficClass,
+                                     uint8_t  const& serviceLevel)
   {
     // Prepare QP attributes
     struct ibv_qp_attr attr = {};
@@ -3300,11 +3306,12 @@ static bool IsConfiguredGid(union ibv_gid const& gid)
       attr.ah_attr.grh.flow_label                = 0;
       attr.ah_attr.grh.sgid_index                = connInfo.gidIdx;
       attr.ah_attr.grh.hop_limit                 = 255;
+      attr.ah_attr.grh.traffic_class             = trafficClass;
     } else {
       attr.ah_attr.is_global = 0;
       attr.ah_attr.dlid      = connInfo.lid;
     }
-    attr.ah_attr.sl            = 0;
+    attr.ah_attr.sl            = serviceLevel;
     attr.ah_attr.src_path_bits = 0;
     attr.ah_attr.port_num      = port;
     attr.dest_qp_num           = connInfo.qpn;
@@ -3588,7 +3595,7 @@ static bool IsConfiguredGid(union ibv_gid const& gid)
       static_assert(std::is_trivially_copyable<QpTransitionResult>::value, "QpTransitionResult must be trivially copyable for MPI broadcast");
       QpTransitionResult srcQpResult = {ERR_NONE, false};
       if (GetRank() == srcMemRank) {
-        ErrResult err = TransitionQpToRtr(rss.srcQueuePairs[i], dstConnInfo, port, srcIsRoCE, rss.srcPortAttr.active_mtu);
+        ErrResult err = TransitionQpToRtr(rss.srcQueuePairs[i], dstConnInfo, port, srcIsRoCE, rss.srcPortAttr.active_mtu, cfg.nic.trafficClass, cfg.nic.serviceLevel);
         srcQpResult.rtrFailed = (err.errType != ERR_NONE);
         if (err.errType == ERR_NONE) {
           err = TransitionQpToRts(rss.srcQueuePairs[i]);
@@ -3603,7 +3610,7 @@ static bool IsConfiguredGid(union ibv_gid const& gid)
 
       QpTransitionResult dstQpResult = {ERR_NONE, false};
       if (GetRank() == dstMemRank) {
-        ErrResult err = TransitionQpToRtr(rss.dstQueuePairs[i], srcConnInfo, port, dstIsRoCE, rss.dstPortAttr.active_mtu);
+        ErrResult err = TransitionQpToRtr(rss.dstQueuePairs[i], srcConnInfo, port, dstIsRoCE, rss.dstPortAttr.active_mtu, cfg.nic.trafficClass, cfg.nic.serviceLevel);
         dstQpResult.rtrFailed = (err.errType != ERR_NONE);
         if (err.errType == ERR_NONE) {
           err = TransitionQpToRts(rss.dstQueuePairs[i]);