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);
+    m_node = VirtualMemory::bindToNUMANode(affinity);
-    Platform::trySetThreadAffinity(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) {
+    if (m_nodes > 1) {
-        m_features |= SET_THISTHREAD_MEMBIND;
+        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];

src/core/Miner.cpp

@@ -366,7 +366,7 @@ void xmrig::Miner::setJob(const Job &job, bool donate)
 
 #   ifdef XMRIG_ALGO_RANDOMX
     const CpuConfig &cpu = d_ptr->controller->config()->cpu();
-    Rx::init(job, cpu.initThreads(), cpu.isHugePages());
+    Rx::init(job, cpu.initThreads(), cpu.isHugePages(), true);
 #   endif
 
     uv_rwlock_wrunlock(&d_ptr->rwlock);

src/crypto/common/VirtualMemory.cpp

@@ -35,11 +35,11 @@
 #include "crypto/common/VirtualMemory.h"
 
 
-void xmrig::VirtualMemory::bindToNUMANode(int64_t affinity)
+uint32_t xmrig::VirtualMemory::bindToNUMANode(int64_t affinity)
 {
 #   ifdef XMRIG_FEATURE_HWLOC
     if (affinity < 0 || !HwlocCpuInfo::has(HwlocCpuInfo::SET_THISTHREAD_MEMBIND)) {
-        return;
+        return 0;
     }
 
     hwloc_topology_t topology;
@@ -53,6 +53,21 @@ void xmrig::VirtualMemory::bindToNUMANode(int64_t affinity)
         LOG_WARN("CPU #%02u warning: \"can't bind memory\"", puId);
     }
 
+    hwloc_obj_t node = nullptr;
+    uint32_t nodeId  = 0;
+
+    while ((node = hwloc_get_next_obj_by_type(topology, HWLOC_OBJ_NUMANODE, node)) != nullptr) {
+      if (hwloc_bitmap_intersects(node->cpuset, pu->cpuset)) {
+          nodeId = node->os_index;
+
+          break;
+      }
+    }
+
     hwloc_topology_destroy(topology);
+
+    return nodeId;
+#   else
+    return 0;
 #   endif
 }

src/crypto/common/VirtualMemory.h

@@ -52,9 +52,9 @@ public:
         return std::pair<size_t, size_t>(isHugePages() ? (align(size()) / 2097152) : 0, align(size()) / 2097152);
     }
 
+    static uint32_t bindToNUMANode(int64_t affinity);
     static void *allocateExecutableMemory(size_t size);
     static void *allocateLargePagesMemory(size_t size);
-    static void bindToNUMANode(int64_t affinity);
     static void flushInstructionCache(void *p, size_t size);
     static void freeLargePagesMemory(void *p, size_t size);
     static void init(bool hugePages);

src/crypto/rx/Rx.cpp

@@ -25,12 +25,20 @@
  */
 
 
+#include <map>
 #include <thread>
 #include <uv.h>
 
 
+#ifdef XMRIG_FEATURE_HWLOC
+#   include <hwloc.h>
+#   include "backend/cpu/platform/HwlocCpuInfo.h"
+#endif
+
+
 #include "backend/cpu/Cpu.h"
 #include "base/io/log/Log.h"
+#include "base/kernel/Platform.h"
 #include "base/net/stratum/Job.h"
 #include "base/tools/Buffer.h"
 #include "base/tools/Chrono.h"
@@ -42,6 +50,9 @@
 namespace xmrig {
 
 
+static const char *tag  = BLUE_BG(WHITE_BOLD_S " rx ") " ";
+
+
 class RxPrivate
 {
 public:
@@ -53,10 +64,12 @@ public:
 
     inline ~RxPrivate()
     {
-        for (RxDataset *dataset : datasets) {
+        for (auto const &item : datasets) {
-            delete dataset;
+            delete item.second;
         }
 
+        datasets.clear();
+
         uv_mutex_destroy(&mutex);
     }
 
@@ -65,23 +78,79 @@ public:
     inline void unlock() { uv_mutex_unlock(&mutex); }
 
 
-    std::vector<RxDataset *> datasets;
+    static void allocate(RxPrivate *self, uint32_t nodeId)
+    {
+        const uint64_t ts = Chrono::steadyMSecs();
+
+#       ifdef XMRIG_FEATURE_HWLOC
+        if (self->numa) {
+            hwloc_topology_t topology;
+            hwloc_topology_init(&topology);
+            hwloc_topology_load(topology);
+
+            hwloc_obj_t node = hwloc_get_numanode_obj_by_os_index(topology, nodeId);
+            if (node) {
+                if (HwlocCpuInfo::has(HwlocCpuInfo::SET_THISTHREAD_MEMBIND)) {
+                    hwloc_set_membind_nodeset(topology, node->nodeset, HWLOC_MEMBIND_BIND, HWLOC_MEMBIND_THREAD);
+                }
+
+                Platform::setThreadAffinity(static_cast<uint64_t>(hwloc_bitmap_first(node->cpuset)));
+            }
+
+            hwloc_topology_destroy(topology);
+        }
+#       endif
+
+        LOG_INFO("%s" CYAN_BOLD("#%u") MAGENTA_BOLD(" allocate") CYAN_BOLD(" %zu MB") BLACK_BOLD(" (%zu+%zu) for RandomX dataset & cache"),
+                 tag,
+                 nodeId,
+                 (RxDataset::size() + RxCache::size()) / 1024 / 1024,
+                 RxDataset::size() / 1024 / 1024,
+                 RxCache::size() / 1024 / 1024
+                 );
+
+        RxDataset *dataset   = new RxDataset(self->hugePages);
+        self->datasets[nodeId] = dataset;
+
+        if (dataset->get() != nullptr) {
+            const auto hugePages = dataset->hugePages();
+            const double percent = hugePages.first == 0 ? 0.0 : static_cast<double>(hugePages.first) / hugePages.second * 100.0;
+
+            LOG_INFO("%s" CYAN_BOLD("#%u") GREEN(" allocate done") " huge pages %s%u/%u %1.0f%%" CLEAR " %sJIT" BLACK_BOLD(" (%" PRIu64 " ms)"),
+                     tag,
+                     nodeId,
+                     (hugePages.first == hugePages.second ? GREEN_BOLD_S : (hugePages.first == 0 ? RED_BOLD_S : YELLOW_BOLD_S)),
+                     hugePages.first,
+                     hugePages.second,
+                     percent,
+                     dataset->cache()->isJIT() ? GREEN_BOLD_S "+" : RED_BOLD_S "-",
+                     Chrono::steadyMSecs() - ts
+                     );
+        }
+        else {
+            LOG_WARN(CLEAR "%s" CYAN_BOLD("#%u") YELLOW_BOLD_S " failed to allocate RandomX dataset, switching to slow mode", tag, nodeId);
+        }
+    }
+
+
+    bool hugePages  = true;
+    bool numa       = true;
+    std::map<uint32_t, RxDataset *> datasets;
     uv_mutex_t mutex;
 };
 
 
 static RxPrivate *d_ptr = new RxPrivate();
-static const char *tag  = BLUE_BG(WHITE_BOLD_S " rx ");
 
 
 } // namespace xmrig
 
 
 
-bool xmrig::Rx::isReady(const Job &job, int64_t)
+bool xmrig::Rx::isReady(const Job &job, uint32_t nodeId)
 {
     d_ptr->lock();
-    const bool rc = isReady(job.seedHash(), job.algorithm());
+    const bool rc = isReady(job.seedHash(), job.algorithm(), d_ptr->numa ? nodeId : 0);
     d_ptr->unlock();
 
     return rc;
@@ -89,38 +158,56 @@ bool xmrig::Rx::isReady(const Job &job, int64_t)
 
 
 
-xmrig::RxDataset *xmrig::Rx::dataset(int64_t)
+xmrig::RxDataset *xmrig::Rx::dataset(uint32_t nodeId)
 {
     d_ptr->lock();
-    RxDataset *dataset = d_ptr->datasets[0];
+    RxDataset *dataset = d_ptr->datasets[d_ptr->numa ? nodeId : 0];
     d_ptr->unlock();
 
     return dataset;
 }
 
 
-void xmrig::Rx::init(const Job &job, int initThreads, bool hugePages)
+void xmrig::Rx::init(const Job &job, int initThreads, bool hugePages, bool numa)
 {
     if (job.algorithm().family() != Algorithm::RANDOM_X) {
         return;
     }
 
     d_ptr->lock();
-    if (d_ptr->datasets.empty()) {
+
-        d_ptr->datasets.push_back(nullptr);
+    size_t ready = 0;
+
+    for (auto const &item : d_ptr->datasets) {
+        if (isReady(job.seedHash(), job.algorithm(), item.first)) {
+            ready++;
+        }
     }
 
-    if (isReady(job.seedHash(), job.algorithm())) {
+    if (!d_ptr->datasets.empty() && ready == d_ptr->datasets.size()) {
         d_ptr->unlock();
 
         return;
     }
 
-    const uint32_t threads  = initThreads < 1 ? static_cast<uint32_t>(Cpu::info()->threads())
+    d_ptr->hugePages       = hugePages;
-                                              : static_cast<uint32_t>(initThreads);
+    d_ptr->numa            = numa && Cpu::info()->nodes() > 1;
+    const uint32_t threads = initThreads < 1 ? static_cast<uint32_t>(Cpu::info()->threads())
+                                             : static_cast<uint32_t>(initThreads);
 
-    std::thread thread(initDataset, 0, job.seedHash(), job.algorithm(), threads, hugePages);
+#   ifdef XMRIG_FEATURE_HWLOC
-    thread.detach();
+    if (d_ptr->numa) {
+        for (uint32_t nodeId : HwlocCpuInfo::nodeIndexes()) {
+            std::thread thread(initDataset, nodeId, job.seedHash(), job.algorithm(), threads);
+            thread.detach();
+        }
+    }
+    else
+#   endif
+    {
+        std::thread thread(initDataset, 0, job.seedHash(), job.algorithm(), threads);
+        thread.detach();
+    }
 
     d_ptr->unlock();
 }
@@ -134,69 +221,56 @@ void xmrig::Rx::stop()
 }
 
 
-bool xmrig::Rx::isReady(const uint8_t *seed, const Algorithm &algorithm)
+bool xmrig::Rx::isReady(const uint8_t *seed, const Algorithm &algorithm, uint32_t nodeId)
 {
-    return !d_ptr->datasets.empty() && d_ptr->datasets[0] != nullptr && d_ptr->datasets[0]->isReady(seed, algorithm);
+    return !d_ptr->datasets.empty() && d_ptr->datasets[nodeId] != nullptr && d_ptr->datasets[nodeId]->isReady(seed, algorithm);
 }
 
 
-void xmrig::Rx::initDataset(size_t index, const uint8_t *seed, const Algorithm &algorithm, uint32_t threads, bool hugePages)
+void xmrig::Rx::initDataset(uint32_t nodeId, const uint8_t *seed, const Algorithm &algorithm, uint32_t threads)
 {
     d_ptr->lock();
 
-    if (!d_ptr->datasets[index]) {
+    RxDataset *dataset = d_ptr->datasets[nodeId];
-        const uint64_t ts = Chrono::steadyMSecs();
 
-        LOG_INFO("%s" MAGENTA_BOLD(" allocate") CYAN_BOLD(" %zu MB") BLACK_BOLD(" (%zu+%zu) for RandomX dataset & cache"),
+    if (!dataset) {
-                 tag,
+#       ifdef XMRIG_FEATURE_HWLOC
-                 (RxDataset::size() + RxCache::size()) / 1024 / 1024,
+        if (d_ptr->numa) {
-                 RxDataset::size() / 1024 / 1024,
+            std::thread thread(RxPrivate::allocate, d_ptr, nodeId);
-                 RxCache::size() / 1024 / 1024
+            thread.join();
-                 );
+        } else
-
+#       endif
-        d_ptr->datasets[index] = new RxDataset(hugePages);
+        {
-
+            RxPrivate::allocate(d_ptr, nodeId);
-        if (d_ptr->datasets[index]->get() != nullptr) {
-            const auto hugePages = d_ptr->datasets[index]->hugePages();
-            const double percent = hugePages.first == 0 ? 0.0 : static_cast<double>(hugePages.first) / hugePages.second * 100.0;
-
-            LOG_INFO("%s" GREEN(" allocate done") " huge pages %s%u/%u %1.0f%%" CLEAR " %sJIT" BLACK_BOLD(" (%" PRIu64 " ms)"),
-                     tag,
-                     (hugePages.first == hugePages.second ? GREEN_BOLD_S : (hugePages.first == 0 ? RED_BOLD_S : YELLOW_BOLD_S)),
-                     hugePages.first,
-                     hugePages.second,
-                     percent,
-                     d_ptr->datasets[index]->cache()->isJIT() ? GREEN_BOLD_S "+" : RED_BOLD_S "-",
-                     Chrono::steadyMSecs() - ts
-                     );
-        }
-        else {
-            LOG_WARN(CLEAR "%s" YELLOW_BOLD_S " failed to allocate RandomX dataset, switching to slow mode", tag);
         }
+
+        dataset = d_ptr->datasets[nodeId];
     }
 
-    if (!d_ptr->datasets[index]->isReady(seed, algorithm)) {
+    if (!dataset->isReady(seed, algorithm)) {
         const uint64_t ts = Chrono::steadyMSecs();
 
-        if (d_ptr->datasets[index]->get() != nullptr) {
+        if (dataset->get() != nullptr) {
-            LOG_INFO("%s" MAGENTA_BOLD(" init dataset") " algo " WHITE_BOLD("%s (") CYAN_BOLD("%u") WHITE_BOLD(" threads)") BLACK_BOLD(" seed %s..."),
+            LOG_INFO("%s" CYAN_BOLD("#%u") MAGENTA_BOLD(" init dataset") " algo " WHITE_BOLD("%s (") CYAN_BOLD("%u") WHITE_BOLD(" threads)") BLACK_BOLD(" seed %s..."),
                      tag,
+                     nodeId,
                      algorithm.shortName(),
                      threads,
                      Buffer::toHex(seed, 8).data()
                      );
         }
         else {
-            LOG_INFO("%s" MAGENTA_BOLD(" init cache") " algo " WHITE_BOLD("%s") BLACK_BOLD(" seed %s..."),
+            LOG_INFO("%s" CYAN_BOLD("#%u") MAGENTA_BOLD(" init cache") " algo " WHITE_BOLD("%s") BLACK_BOLD(" seed %s..."),
                      tag,
+                     nodeId,
                      algorithm.shortName(),
                      Buffer::toHex(seed, 8).data()
                      );
         }
 
-        d_ptr->datasets[index]->init(seed, algorithm, threads);
+        dataset->init(seed, algorithm, threads);
 
-        LOG_INFO("%s" GREEN(" init done") BLACK_BOLD(" (%" PRIu64 " ms)"), tag, Chrono::steadyMSecs() - ts);
+        LOG_INFO("%s" CYAN_BOLD("#%u") GREEN(" init done") BLACK_BOLD(" (%" PRIu64 " ms)"), tag, nodeId, Chrono::steadyMSecs() - ts);
     }
 
     d_ptr->unlock();

src/crypto/rx/Rx.h

@@ -43,14 +43,14 @@ class Job;
 class Rx
 {
 public:
-    static bool isReady(const Job &job, int64_t affinity);
+    static bool isReady(const Job &job, uint32_t nodeId);
-    static RxDataset *dataset(int64_t affinity);
+    static RxDataset *dataset(uint32_t nodeId);
-    static void init(const Job &job, int initThreads, bool hugePages);
+    static void init(const Job &job, int initThreads, bool hugePages, bool numa);
     static void stop();
 
 private:
-    static bool isReady(const uint8_t *seed, const Algorithm &algorithm);
+    static bool isReady(const uint8_t *seed, const Algorithm &algorithm, uint32_t nodeId);
-    static void initDataset(size_t index, const uint8_t *seed, const Algorithm &algorithm, uint32_t threads, bool hugePages);
+    static void initDataset(uint32_t nodeId, const uint8_t *seed, const Algorithm &algorithm, uint32_t threads);
 };