Add backend override config to ctranComm init

Summary:
As title.
1. Add backend field to the ctranConfig struct.
2. Allow ctranConfig to be passed when ctranComm is created.

Reviewed By: saifhhasan

Differential Revision: D85444909

fbshipit-source-id: 89c98ad7a4b7b39e4b81fa12ae794ade4b32f195

Author: function47

comms/ctran/CtranComm.h

@@ -16,16 +16,19 @@
 #include "comms/utils/colltrace/CollTraceInterface.h"
 #include "comms/utils/commSpecs.h"
 
+using meta::comms::CommBackend;
 struct ctranConfig {
   int blocking{-1};
   const char* commDesc{nullptr};
   const char* ncclAllGatherAlgo{nullptr};
+  std::vector<enum CommBackend> backends = {};
 
   bool operator==(const ctranConfig& other) const {
     return (
         blocking == other.blocking && commDesc == other.commDesc &&
         commDesc == other.commDesc &&
-        ncclAllGatherAlgo == other.ncclAllGatherAlgo);
+        ncclAllGatherAlgo == other.ncclAllGatherAlgo &&
+        backends == other.backends);
   }
 };
 
@@ -46,8 +49,9 @@ class CtranComm {
   // For real communicationator we should use factory method to create.
   explicit CtranComm(
       std::shared_ptr<Abort> abort =
-          ctran::utils::createAbort(/*enabled=*/false))
-      : abort_(abort) {
+          ctran::utils::createAbort(/*enabled=*/false),
+      ctranConfig commConfig = ctranConfig{})
+      : config_(commConfig), abort_(abort) {
     asyncErr_ =
         std::make_shared<AsyncError>(NCCL_CTRAN_ABORT_ON_ERROR, "CtranComm");
     if (!abort_) {
@@ -58,11 +62,12 @@ class CtranComm {
     opCount_ = &ctranOpCount_;
   }
 
-  // The MemCache allocator is destroyed in a different time than all other
-  // Ctran resources. To accommodate this, we split the CtranComm destructor
-  // into two parts. In the first part, we destroy all resources except for
-  // MemCache. The second part is moved to the destructor, where it is safe to
-  // destroy MemCache and reset its reference.
+  // The MemCache allocator is destroyed in a different time than all
+  // other Ctran resources. To accommodate this, we split the CtranComm
+  // destructor into two parts. In the first part, we destroy all
+  // resources except for MemCache. The second part is moved to the
+  // destructor, where it is safe to destroy MemCache and reset its
+  // reference.
   void destroy() {
     // All smart pointers are automatically de-initialized, but we want to
     // ensure they do so in a specific order. Therefore, we manually handle
@@ -72,8 +77,8 @@ class CtranComm {
     collTrace_.reset();
     colltraceNew_.reset();
     statex_.reset();
-    // NOTE: memCache needs to be destroyed after transportProxy_ to release all
-    // buffers
+    // NOTE: memCache needs to be destroyed after transportProxy_ to release
+    // all buffers
     memCache_.reset();
 
     this->logMetaData_.commDesc.clear();

comms/ctran/tests/CtranCommTest.cc

@@ -49,4 +49,17 @@ TEST(CtranCommTest, AbortAvailableAndDisabled) {
   EXPECT_FALSE(comm.testAbort());
 }
 
+TEST(CtranCommTest, ctranCommConfigTest) {
+  auto abort = ctran::utils::createAbort(/*enabled=*/true);
+  ctranConfig config = {
+      .backends = {CommBackend::IB, CommBackend::NVL, CommBackend::SOCKET}};
+
+  CtranComm comm(abort, config);
+  EXPECT_EQ(comm.config_.backends.size(), 3);
+
+  /// Explictly create comm with false abort as first argument is unommitable
+  CtranComm comm2(ctran::utils::createAbort(false));
+  EXPECT_EQ(comm2.config_.backends.size(), 0);
+}
+
 } // namespace ctran::testing

comms/utils/commSpecs.h

@@ -299,6 +299,15 @@ enum class CommProtocol {
   NumProtocols = 3 // Simple/LL/LL128
 };
 
+enum class CommBackend {
+  UNSET = 0,
+  IB = 1,
+  NVL = 2,
+  SOCKET = 3,
+  TCPDM = 4,
+  NUM_BACKENDS = 5
+};
+
 class CommsError {
  public:
   CommsError(std::string msg, commResult_t code)