Added support for allocate RandomX dataset on each NUMA node.

src/backend/common/Worker.cpp

@@ -37,9 +37,9 @@ xmrig::Worker::Worker(size_t id, int64_t affinity, int priority) :
     m_timestamp(0),
     m_count(0)
 {
-    VirtualMemory::bindToNUMANode(affinity);
-    Platform::trySetThreadAffinity(affinity);
+    m_node = VirtualMemory::bindToNUMANode(affinity);
 
+    Platform::trySetThreadAffinity(affinity);
     Platform::setThreadPriority(priority);
 }
 

src/backend/common/Worker.h

@@ -54,6 +54,7 @@ protected:
     const size_t m_id;
     std::atomic<uint64_t> m_hashCount;
     std::atomic<uint64_t> m_timestamp;
+    uint32_t m_node = 0;
     uint64_t m_count;
 };
 

src/backend/cpu/CpuBackend.cpp

@@ -295,7 +295,7 @@ rapidjson::Value xmrig::CpuBackend::toJSON(rapidjson::Document &doc) const
 
 #   ifdef XMRIG_ALGO_RANDOMX
     if (d_ptr->algo.family() == Algorithm::RANDOM_X) {
-        RxDataset *dataset = Rx::dataset(-1); // FIXME
+        RxDataset *dataset = Rx::dataset(0); // FIXME
         if (dataset) {
             const auto rxPages = dataset->hugePages();
             pages[0] += rxPages.first;

src/backend/cpu/CpuWorker.cpp

@@ -24,6 +24,7 @@
  */
 
 
+#include <assert.h>
 #include <thread>
 
 
@@ -81,15 +82,20 @@ xmrig::CpuWorker<N>::~CpuWorker()
 template<size_t N>
 void xmrig::CpuWorker<N>::allocateRandomX_VM()
 {
-    while (!Rx::isReady(m_job.currentJob(), m_affinity)) {
+    while (!Rx::isReady(m_job.currentJob(), m_node)) {
         std::this_thread::sleep_for(std::chrono::milliseconds(200));
 
         if (Nonce::sequence(Nonce::CPU) == 0) {
-            break;
+            return;
         }
     }
 
-    RxDataset *dataset = Rx::dataset(m_affinity);
+    RxDataset *dataset = Rx::dataset(m_node);
+    assert(dataset != nullptr);
+
+    if (!dataset) {
+        return;
+    }
 
     if (!m_vm) {
         m_vm = new RxVm(dataset, m_memory->scratchpad(), !m_hwAES);

src/backend/cpu/platform/HwlocCpuInfo.cpp

@@ -38,6 +38,7 @@
 namespace xmrig {
 
 
+std::vector<uint32_t> HwlocCpuInfo::m_nodeIndexes;
 uint32_t HwlocCpuInfo::m_features = 0;
 
 
@@ -151,8 +152,17 @@ xmrig::HwlocCpuInfo::HwlocCpuInfo() : BasicCpuInfo(),
     m_nodes     = std::max<size_t>(countByType(m_topology, HWLOC_OBJ_NUMANODE), 1);
     m_packages  = countByType(m_topology, HWLOC_OBJ_PACKAGE);
 
-    if (nodes() > 1 && hwloc_topology_get_support(m_topology)->membind->set_thisthread_membind) {
-        m_features |= SET_THISTHREAD_MEMBIND;
+    if (m_nodes > 1) {
+        if (hwloc_topology_get_support(m_topology)->membind->set_thisthread_membind) {
+            m_features |= SET_THISTHREAD_MEMBIND;
+        }
+
+        m_nodeIndexes.reserve(m_nodes);
+        hwloc_obj_t node = nullptr;
+
+        while ((node = hwloc_get_next_obj_by_type(m_topology, HWLOC_OBJ_NUMANODE, node)) != nullptr) {
+            m_nodeIndexes.emplace_back(node->os_index);
+        }
     }
 }
 

src/backend/cpu/platform/HwlocCpuInfo.h

@@ -47,7 +47,8 @@ public:
     HwlocCpuInfo();
     ~HwlocCpuInfo() override;
 
-    static inline bool has(Feature feature) { return m_features & feature; }
+    static inline bool has(Feature feature)                     { return m_features & feature; }
+    static inline const std::vector<uint32_t> &nodeIndexes()    { return m_nodeIndexes; }
 
 protected:
     CpuThreads threads(const Algorithm &algorithm) const override;
@@ -62,6 +63,7 @@ protected:
 private:
     void processTopLevelCache(hwloc_obj_t obj, const Algorithm &algorithm, CpuThreads &threads) const;
 
+    static std::vector<uint32_t> m_nodeIndexes;
     static uint32_t m_features;
 
     char m_backend[20];