hwloc NUMA memory binding

- add function `bindMemoryToNUMANode()`
- add `affinity` member to `minethd`

hwlocMemory.hpp

@@ -0,0 +1,53 @@
+#pragma once
+
+#include "console.h"
+
+#ifndef CONF_NO_HWLOC
+
+#include <hwloc.h>
+
+/** pin memory to NUMA node
+ *
+ * Set the default memory policy for the current thread to bind memory to the
+ * NUMA node.
+ *
+ * @param puId core id
+ */
+void bindMemoryToNUMANode( int puId )
+{
+	int depth;
+	hwloc_topology_t topology;
+	hwloc_obj_t obj;
+
+	hwloc_topology_init(&topology);
+	hwloc_topology_load(topology);
+
+	depth = hwloc_get_type_depth(topology, HWLOC_OBJ_PU);
+
+	for( int i = 0;
+		i < hwloc_get_nbobjs_by_depth(topology, depth);
+		i++ )
+	{
+		hwloc_obj_t pu = hwloc_get_obj_by_depth(topology, depth, i);
+		if(  pu->os_index == puId )
+			if( 0 > hwloc_set_membind_nodeset(
+				topology,
+				pu->nodeset,
+				HWLOC_MEMBIND_BIND,
+				HWLOC_MEMBIND_THREAD)
+			)
+				printer::inst()->print_msg(L0, "hwloc: can't bind memory");
+			else
+			{
+				printer::inst()->print_msg(L0, "hwloc: memory pinned");
+				break;
+			}
+	}
+}
+#else
+
+void bindMemoryToNUMANode( int )
+{
+}
+
+#endif

minethd.cpp

@@ -64,6 +64,7 @@ void thd_setaffinity(std::thread::native_handle_type h, uint64_t cpu_id)
 #include "minethd.h"
 #include "jconf.h"
 #include "crypto/cryptonight_aesni.h"
+#include "hwlocMemory.hpp"
 
 telemetry::telemetry(size_t iThd)
 {
@@ -140,7 +141,7 @@ void telemetry::push_perf_value(size_t iThd, uint64_t iHashCount, uint64_t iTime
 	iBucketTop[iThd] = (iTop + 1) & iBucketMask;
 }
 
-minethd::minethd(miner_work& pWork, size_t iNo, bool double_work, bool no_prefetch)
+minethd::minethd(miner_work& pWork, size_t iNo, bool double_work, bool no_prefetch, int affinity)
 {
 	oWork = pWork;
 	bQuit = 0;
@@ -149,6 +150,7 @@ minethd::minethd(miner_work& pWork, size_t iNo, bool double_work, bool no_prefet
 	iHashCount = 0;
 	iTimestamp = 0;
 	bNoPrefetch = no_prefetch;
+	this->affinity = affinity;
 
 	if(double_work)
 		oWorkThd = std::thread(&minethd::double_work_main, this);
@@ -295,7 +297,7 @@ std::vector<minethd*>* minethd::thread_starter(miner_work& pWork)
 	{
 		jconf::inst()->GetThreadConfig(i, cfg);
 
-		minethd* thd = new minethd(pWork, i, cfg.bDoubleMode, cfg.bNoPrefetch);
+		minethd* thd = new minethd(pWork, i, cfg.bDoubleMode, cfg.bNoPrefetch, cfg.iCpuAff);
 
 		if(cfg.iCpuAff >= 0)
 		{
@@ -361,6 +363,9 @@ minethd::cn_hash_fun minethd::func_selector(bool bHaveAes, bool bNoPrefetch)
 
 void minethd::work_main()
 {
+	// pin memory to NUMA node
+	bindMemoryToNUMANode(this->affinity);
+
 	cn_hash_fun hash_fun;
 	cryptonight_ctx* ctx;
 	uint64_t iCount = 0;
@@ -448,6 +453,9 @@ minethd::cn_hash_fun_dbl minethd::func_dbl_selector(bool bHaveAes, bool bNoPrefe
 
 void minethd::double_work_main()
 {
+	// pin memory to NUMA node
+	bindMemoryToNUMANode(this->affinity);
+
 	cn_hash_fun_dbl hash_fun;
 	cryptonight_ctx* ctx0;
 	cryptonight_ctx* ctx1;

minethd.h

@@ -97,11 +97,13 @@ public:
 	std::atomic<uint64_t> iHashCount;
 	std::atomic<uint64_t> iTimestamp;
 
+	int affinity;
+
 private:
 	typedef void (*cn_hash_fun)(const void*, size_t, void*, cryptonight_ctx*);
 	typedef void (*cn_hash_fun_dbl)(const void*, size_t, void*, cryptonight_ctx* __restrict, cryptonight_ctx* __restrict);
 
-	minethd(miner_work& pWork, size_t iNo, bool double_work, bool no_prefetch);
+	minethd(miner_work& pWork, size_t iNo, bool double_work, bool no_prefetch, int affinity);
 
 	// We use the top 10 bits of the nonce for thread and resume
 	// This allows us to resume up to 128 threads 4 times before