Merge branch 'main' into rust_dl_apis_serialization

Author: evgeny-leksikov

.ci/jenkins/lib/build-container-matrix.yaml

@@ -11,7 +11,7 @@ timeout_minutes: 240
 # Infrastructure
 kubernetes:
   cloud: il-ipp-blossom-prod
-  namespace: swx-media
+  namespace: nbu-swx-nixl
   limits: "{memory: 16Gi, cpu: 8000m}"
   requests: "{memory: 8Gi, cpu: 4000m}"
 
@@ -93,6 +93,7 @@ steps:
 
   - name: Add Version Info
     run: |
+      git config --global --add safe.directory '*'
       # Extract standardized 8-char commit hash for UCX version info:
       if [[ "$BUILD_TARGET" == "nixlbench" ]]; then
         CLEAN_UCX=$(cd "$WORKSPACE/ucx-src" && git rev-parse --short=8 HEAD)

.ci/jenkins/lib/build-matrix.yaml

@@ -29,7 +29,7 @@ timeout_minutes: 240
 
 kubernetes:
   cloud: il-ipp-blossom-prod
-  namespace: swx-media
+  namespace: nbu-swx-nixl
   limits: "{memory: 8Gi, cpu: 8000m}"
   requests: "{memory: 8Gi, cpu: 8000m}"
 

.ci/jenkins/pipeline/proj-jjb.yaml

@@ -9,7 +9,6 @@
 - job-template:
     name: "{jjb_proj}-dispatcher"  # Will be expanded to 'nixl-ci-dispatcher'
     project-type: pipeline
-    folder: "{jjb_folder}"
     properties:
         # GitHub integration settings
         - github:
@@ -56,49 +55,39 @@
           githubHelper = GithubHelper.getInstance("${{GIT_PASSWORD}}", VARIABLE_FROM_POST)
         }}
         // Update GitHub commit status
-        githubHelper.updateCommitStatus("$BUILD_URL", "CI Dispatcher started", GitHubCommitState.PENDING)
+        def blueOceanUrl = "${{JENKINS_URL}}blue/organizations/jenkins/nixl-ci-dispatcher/detail/nixl-ci-dispatcher/${{BUILD_NUMBER}}/pipeline/"
+        githubHelper.updateCommitStatus(blueOceanUrl, "NIXL CI Started", GitHubCommitState.PENDING)
         currentBuild.description = githubHelper.getBuildDescription()
         try {{
             // Trigger the actual build and test jobs in parallel
             parallel build: {{
                 def buildJob = 'nixl-ci-build'
-                def build = build job: buildJob, parameters: [
+                build job: buildJob, parameters: [
                     string(name: 'sha1', value: githubHelper.getMergedSHA()),
                     string(name: 'githubData', value: VARIABLE_FROM_POST)
-                ], propagate: false
-                currentBuild.description += "<br>Job: <a href='${{JENKINS_URL}}blue/organizations/jenkins/NIXL%2F${{buildJob}}/detail/${{buildJob}}/${{build.number}}/pipeline/'>${{buildJob}}</a> Result: <b style='color:${{build.result == 'SUCCESS' ? 'green' : 'red'}}'>${{build.result}}</b>"
-                if (!build.resultIsBetterOrEqualTo('SUCCESS')) {{
-                  currentBuild.result = build.result
-                  error("Job ${{buildJob}} failed")
-                }}
+                ]
             }}, test: {{
                 def buildJob = 'nixl-ci-test'
-                def build = build job: buildJob, parameters: [
+                build job: buildJob, parameters: [
                     string(name: 'sha1', value: githubHelper.getMergedSHA()),
                     string(name: 'githubData', value: VARIABLE_FROM_POST)
-                ], propagate: false
-                currentBuild.description += "<br>Job: <a href='${{JENKINS_URL}}blue/organizations/jenkins/NIXL%2F${{buildJob}}/detail/${{buildJob}}/${{build.number}}/pipeline/'>${{buildJob}}</a> Result: <b style='color:${{build.result == 'SUCCESS' ? 'green' : 'red'}}'>${{build.result}}</b>"
-                if (!build.resultIsBetterOrEqualTo('SUCCESS')) {{
-                  currentBuild.result = build.result
-                  error("Job ${{buildJob}} failed")
-                }}
+                ]
             }},
             failFast: false  // Continue even if some parallel jobs fail
 
-            githubHelper.updateCommitStatus("$BUILD_URL", "CI Dispatcher successeded", GitHubCommitState.SUCCESS)
+            githubHelper.updateCommitStatus(blueOceanUrl, "NIXL CI Succeeded", GitHubCommitState.SUCCESS)
         }} catch(Exception ex) {{
             // Handle build failures
             currentBuild.result = 'FAILURE'
             println ex
-            githubHelper.updateCommitStatus("$BUILD_URL", "CI Dispatcher failed", GitHubCommitState.FAILURE)
+            githubHelper.updateCommitStatus(blueOceanUrl, "NIXL CI Failed", GitHubCommitState.FAILURE)
             error("failed")
         }}
 
 # Template for the main build job that performs the actual build process
 - job-template:
     name: "{jjb_proj}-build"      # Will be expanded to 'nixl-ci-build'
     project-type: pipeline
-    folder: "{jjb_folder}"
     disabled: false
     properties:
         # Similar properties as dispatcher job
@@ -161,7 +150,6 @@
 - job-template:
     name: "{jjb_proj}-test"      # Will be expanded to 'nixl-ci-test'
     project-type: pipeline
-    folder: "{jjb_folder}"
     disabled: false
     properties:
         # Similar properties as dispatcher job
@@ -226,7 +214,6 @@
 - job-template:
     name: "{jjb_proj}-build-container"
     project-type: pipeline
-    folder: "{jjb_folder}"
     disabled: false
     properties:
       - build-discarder:
@@ -325,7 +312,6 @@
     jjb_proj: 'nixl-ci'           # Project prefix for job names
     jjb_git: 'git@github.com:ai-dynamo/nixl.git'  # Repository URL
     jjb_jenkinsfile: '.ci/jenkins/pipeline/Jenkinsfile'  # Main pipeline definition
-    jjb_folder: 'NIXL'
     jjb_branch: 'main'            # Default branch
     jjb_gh_url: 'https://github.com/ai-dynamo/nixl'  # GitHub web URL
     jobs:

benchmark/nixlbench/src/worker/nixl/nixl_worker.cpp

@@ -862,7 +862,7 @@ xferBenchNixlWorker::allocateMemory(int num_threads) {
         iov_lists.push_back(iov_list);
 
         /* Workaround for a GUSLI registration bug which resets memory to 0 */
-        if (seg_type == DRAM_SEG) {
+        if (XFERBENCH_BACKEND_GUSLI == xferBenchConfig::backend && seg_type == DRAM_SEG) {
             for (auto &iov : iov_list) {
                 if (isInitiator()) {
                     memset((void *)iov.addr, XFERBENCH_INITIATOR_BUFFER_ELEMENT, buffer_size);

docs/BackendGuide.md

@@ -108,16 +108,16 @@ Note that getNotif does not know which agent it should look for to receive the n
 
 A key underlying abstraction for NIXL library is a descriptor list, that is made of a memory space (host/GPU/block/File/Obj-Store) and a list of descriptors. There are 2 types of descriptors used for the SB API.
 
-*For transfers: (addr, len, devID, metadata), where metadata is a pointer to an nixlBackendMD object relevant to the registered memory that this descriptor falls within.
-*For registration, (addr, len, devID, str) where str is an optional byte-array for extra information. The table below shows the meaning of devID for different memory spaces, as well as optional meaning for File and Object-Store.
+* For transfers: (addr, len, devID, metadata), where metadata is a pointer to an nixlBackendMD object relevant to the registered memory that this descriptor falls within.
+* For registration, (addr, len, devID, str) where str is an optional byte-array for extra information. The table below shows the meaning of devID for different memory spaces, as well as optional meaning for File and Object-Store.
 
 | mem type | addr   | len  | devID         | str (byte-array)           |
 | -------- | ------ | ---- | ------------- | -------------------------- |
 | DRAM     |        |      | 0 (or region) |        -                   |
 | VRAM     |        |      | GPU ID        |        -                   |
 | BLK      |        |      | Vol ID        |        -                   |
 | FILE     | offset | Or 0 | fd            | Path + (access mode)       |
-| DRAM     | offset | Or 0 | key           | Extended key (+ bucket ID) |
+| OBJ      | offset | Or 0 | key           | Extended key (+ bucket ID) |
 
 ## Plugin Manager API
 
@@ -165,7 +165,7 @@ In this step, if the plugin supports talking to remote agents, the required conn
 
 When a connection is requested to an remote agent, which is possible if the remote agent’s metadata is already loaded, the local Agent would look for common backend plugins between itself and the remote agent, and for each of them initiate a connect by using the **connect** API in SB API of such backends.
 
-### Register (Degister) memory with NIXL:
+### Register (Deregister) memory with NIXL:
 
 The agent will receive a list of allocated memories and desired backend from the user, and then will give only one element at a time to the specified backend. Note that backends usually require to register the memories they will access during transfers, and based on that registration keep some metadata for that memory region. For instance, in case of UCX, per each contiguous region of memory, it will produce some local metadata for that region. Agent will give only a single contiguous region of memory to the **register** call in SB API, and in return gets a key (a pointer) to the metadata that backend created for this memory region. Later on, during transfer, the agent will give the same key back to the backend, so backends do not need to do any bookkeeping of such metadata.
 

src/infra/nixl_descriptors.cpp

@@ -112,19 +112,19 @@ nixlBlobDesc::nixlBlobDesc(const nixlBasicDesc &desc,
 }
 
 nixlBlobDesc::nixlBlobDesc(const nixl_blob_t &blob) {
-    size_t meta_size = blob.size() - sizeof(nixlBasicDesc);
-    if (meta_size > 0) {
+    if (blob.size() > sizeof(nixlBasicDesc)) {
+        const size_t meta_size = blob.size() - sizeof(nixlBasicDesc);
         metaInfo.resize(meta_size);
         blob.copy(reinterpret_cast<char*>(this), sizeof(nixlBasicDesc));
         blob.copy(reinterpret_cast<char*>(&metaInfo[0]),
                  meta_size, sizeof(nixlBasicDesc));
-    } else if (meta_size == 0) {
+    } else if (blob.size() == sizeof(nixlBasicDesc)) {
         blob.copy(reinterpret_cast<char*>(this), sizeof(nixlBasicDesc));
     } else { // Error
         addr  = 0;
         len   = 0;
         devId = 0;
-        metaInfo.resize(0);
+        metaInfo.clear();
     }
 }
 

test/unit/plugins/ucx/ucx_backend_test.cpp

@@ -274,24 +274,6 @@ void *releaseValidationPtr(nixl_mem_t mem_type, void *addr)
     return nullptr;
 }
 
-void
-allocateWrongGPUTest(nixlUcxEngine *ucx, int dev_id) {
-    nixlBlobDesc desc = {0};
-    nixlBackendMD* md;
-    void* buf;
-
-    allocateBuffer(VRAM_SEG, dev_id, desc.len, buf);
-
-    desc.devId = dev_id;
-    desc.addr = (uint64_t) buf;
-
-    int ret = ucx->registerMem(desc, VRAM_SEG, md);
-
-    nixl_exit_on_failure((ret == NIXL_ERR_NOT_SUPPORTED), "Failed to register memory", "test");
-
-    releaseBuffer(VRAM_SEG, dev_id, buf);
-}
-
 void
 allocateAndRegister(nixlUcxEngine *ucx,
                     int dev_id,
@@ -749,14 +731,6 @@ int main()
 #endif
     }
 
-#ifdef HAVE_CUDA
-    if (n_vram_dev > 1) {
-		//Test if registering on a different GPU fails correctly
-		allocateWrongGPUTest(ucx[0][0], 1);
-		std::cout << "Verified registration on wrong GPU fails correctly\n";
-	}
-#endif
-
     // Deallocate UCX engines
     for(int i = 0; i < 2; i++) {
         for(int j = 0; j < 2; j++) {