RFR (S/M): 8213827: NUMA heap allocation does not respect process membind/interleave settings [Was: Re: [PATCH] JDK NUMA Interleaving issue]

amith pawar amith.pawar at gmail.com
Sat Jan 12 17:27:44 UTC 2019


Hi Thomas,

SPECJBB shows following improvements with latest patch.
1. max-jOPs around 7-9%
2. critical-jOPS around 4-50%

In webrev.4, Sangheon suggested following change is missing
+      ls.print("UseNUMA is enabled and invoked in '%s' mode."
+                " Heap will be configured using NUMA memory nodes:",
numa_mode);
There is one more space before " Heap.... ", please remove it.

Also os_linux.hpp is already updated for new copyright year so patch import
fails.

The attached patch contains these changes. Please do check.

Thanks,
Amit

On Fri, Jan 11, 2019 at 7:58 PM Thomas Schatzl <thomas.schatzl at oracle.com>
wrote:

> Hi all,
>
>   I prepared new webrevs with the suggestions from Sangheon, and I
> think the changes Amit did:
>
> http://cr.openjdk.java.net/~tschatzl/8213827/webrev.3_to_4/ (diff)
> http://cr.openjdk.java.net/~tschatzl/8213827/webrev.4/ (full)
>
> There were additional fixes:
> - crash at startup with -XX:+UseNUMA since numa_interleave_memory() is
> called before NUMA support has been initialized due to some earlier
> refactoring (the cause is wrong library initialization order at startup
> imho, but is to be handled separately).
> - also CamelCased the enum values.
> - removed some too compliated code ("return (a == b) ? true : false;"
> and one other).
> - removed some more random spacing
>
> This patch passes hs-tier1-5 without additional failures now. Some
> manual testing showed that it seems to do the right thing too now.
>
> @Amith: please check if the change still works on your applications and
> still gives the expected performance improvements.
>
> Thanks,
>   Thomas
>
> On Thu, 2019-01-10 at 16:48 +0530, amith pawar wrote:
> > Hi Sangheon,
> >
> > Thanks again. I have done the required changes and created webrev.
> > Please use following link to download the same as gmail is not
> > allowing to attach.
> >
> > https://drive.google.com/open?id=1QzmW6LdmKbBNHp4-hlcIr9DKY7anUMBy
> >
> > Thanks,
> > Amit
> >
> > On Thu, Jan 10, 2019 at 1:03 AM <sangheon.kim at oracle.com> wrote:
> > > Hi Amith,
> > >
> > > On 1/9/19 3:30 AM, amith pawar wrote:
> > > > Hi Sangheon,
> > > >
> > > > Thanks for reviewing and updated with suggested changes. please
> > > > check.
> > >  Thank you for addressing my comments.
> > > But I can't see below comments addressed:
> > > > > - Looking at 'enum' at os.hpp, we use Camel style.
> > >  I meant to change from 'Numa_allocation_policy' to
> > > 'NumaAllocationPolicy'.
> > > > > - As usual, copyright year updates. I know it was correct when
> > > > > you posted. :)
> > >  Looking at the latest source code, only os_linux.hpp needs a new
> > > copyright year.
> > > - * Copyright (c) 1999, 2018, Oracle and/or its affiliates. All
> > > rights reserved.
> > > + * Copyright (c) 1999, 2019, Oracle and/or its affiliates. All
> > > rights reserved.
> > >
> > > Looking at the v5,
> > > +      ls.print("UseNUMA is enabled and invoked in '%s' mode."
> > > +                " Heap will be configured using NUMA memory
> > > nodes:", numa_mode);
> > > There is one more space before " Heap.... ", please remove it.
> > >
> > > I see the latest version that Thomas posted is v3, but your
> > > attached version is v5. :)
> > >
> > > In addition, it would be better to provide webrev instead of a
> > > patch. ( http://openjdk.java.net/guide/codeReview.html )
> > >
> > > Thanks,
> > > Sangheon
> > >
> > > > Thanks,
> > > > Amit Pawar
> > > >
> > > > On Wed, Jan 9, 2019 at 12:45 AM <sangheon.kim at oracle.com> wrote:
> > > > > Hi Thomas,
> > > > >
> > > > > On 12/13/18 2:33 AM, Thomas Schatzl wrote:
> > > > > > Hi Amit,
> > > > > > On Thu, 2018-12-13 at 15:11 +0530, amith pawar wrote:
> > > > > > > Hi Thomas,
> > > > > > >
> > > > > > > Please find the attached patch updated as per your
> > > > > > > suggestion.
> > > > > > > If everything OK then can you please commit this to repo ?
> > > > > >
> > > > > >   looks good. We will need a second reviewer though, I am
> > > > > > going to ask
> > > > > > around.
> > > > > >
> > > > > > Latest webrev:
> > > > > > http://cr.openjdk.java.net/~tschatzl/8213827/webrev.3/
> > > > >  Webrev.3 looks good to me.
> > > > >
> > > > > I have some minor nits:
> > > > > ----------------------------------------
> > > > > src/hotspot/os/linux/os_linux.cpp
> > > > > 5012       for (int node = 0; node < Linux::numa_max_node();
> > > > > node++) {
> > > > > - Looks like 'node <= Linux::numa_max_node()' is the right one
> > > > > to print the latest node?
> > > > >
> > > > > ----------------------------------------
> > > > > src/hotspot/os/linux/os_linux.hpp
> > > > >  271   enum Numa_allocation_policy{
> > > > > - Looking at 'enum' at os.hpp, we use Camel style.
> > > > > - There are missing space before '{'.
> > > > >
> > > > > - As usual, copyright year updates. I know it was correct when
> > > > > you posted. :)
> > > > >
> > > > > Thanks,
> > > > > Sangheon
> > > > >
> > > > >
> > > > > > Thanks,
> > > > > >   Thomas
> > > > > >
> > > > > >
> > > > >
> > > > >
> > > >
> > > >
> > > > --
> > > > With best regards,
> > > > amit pawar
> > >
> >
> >
>
>
>

-- 
With best regards,
amit pawar
-------------- next part --------------
diff -r 5d7e4d832868 src/hotspot/os/linux/os_linux.cpp
--- a/src/hotspot/os/linux/os_linux.cpp	Sat Jan 12 13:33:18 2019 +0100
+++ b/src/hotspot/os/linux/os_linux.cpp	Sat Jan 12 23:11:54 2019 +0530
@@ -33,6 +33,7 @@
 #include "compiler/disassembler.hpp"
 #include "interpreter/interpreter.hpp"
 #include "logging/log.hpp"
+#include "logging/logStream.hpp"
 #include "memory/allocation.inline.hpp"
 #include "memory/filemap.hpp"
 #include "oops/oop.inline.hpp"
@@ -2780,7 +2781,7 @@
 
   // Get the total number of nodes in the system including nodes without memory.
   for (node = 0; node <= highest_node_number; node++) {
-    if (isnode_in_existing_nodes(node)) {
+    if (is_node_in_existing_nodes(node)) {
       num_nodes++;
     }
   }
@@ -2796,7 +2797,7 @@
   // node number. If the nodes have been bound explicitly using numactl membind,
   // then allocate memory from those nodes only.
   for (int node = 0; node <= highest_node_number; node++) {
-    if (Linux::isnode_in_bound_nodes((unsigned int)node)) {
+    if (Linux::is_node_in_bound_nodes((unsigned int)node)) {
       ids[i++] = node;
     }
   }
@@ -2899,11 +2900,15 @@
                                        libnuma_dlsym(handle, "numa_distance")));
       set_numa_get_membind(CAST_TO_FN_PTR(numa_get_membind_func_t,
                                           libnuma_v2_dlsym(handle, "numa_get_membind")));
+      set_numa_get_interleave_mask(CAST_TO_FN_PTR(numa_get_interleave_mask_func_t,
+                                                  libnuma_v2_dlsym(handle, "numa_get_interleave_mask")));
 
       if (numa_available() != -1) {
         set_numa_all_nodes((unsigned long*)libnuma_dlsym(handle, "numa_all_nodes"));
         set_numa_all_nodes_ptr((struct bitmask **)libnuma_dlsym(handle, "numa_all_nodes_ptr"));
         set_numa_nodes_ptr((struct bitmask **)libnuma_dlsym(handle, "numa_nodes_ptr"));
+        set_numa_interleave_bitmask(_numa_get_interleave_mask());
+        set_numa_membind_bitmask(_numa_get_membind());
         // Create an index -> node mapping, since nodes are not always consecutive
         _nindex_to_node = new (ResourceObj::C_HEAP, mtInternal) GrowableArray<int>(0, true);
         rebuild_nindex_to_node_map();
@@ -2929,7 +2934,7 @@
 
   nindex_to_node()->clear();
   for (int node = 0; node <= highest_node_number; node++) {
-    if (Linux::isnode_in_existing_nodes(node)) {
+    if (Linux::is_node_in_existing_nodes(node)) {
       nindex_to_node()->append(node);
     }
   }
@@ -2966,16 +2971,16 @@
     // the closest configured node. Check also if node is bound, i.e. it's allowed
     // to allocate memory from the node. If it's not allowed, map cpus in that node
     // to the closest node from which memory allocation is allowed.
-    if (!isnode_in_configured_nodes(nindex_to_node()->at(i)) ||
-        !isnode_in_bound_nodes(nindex_to_node()->at(i))) {
+    if (!is_node_in_configured_nodes(nindex_to_node()->at(i)) ||
+        !is_node_in_bound_nodes(nindex_to_node()->at(i))) {
       closest_distance = INT_MAX;
       // Check distance from all remaining nodes in the system. Ignore distance
       // from itself, from another non-configured node, and from another non-bound
       // node.
       for (size_t m = 0; m < node_num; m++) {
         if (m != i &&
-            isnode_in_configured_nodes(nindex_to_node()->at(m)) &&
-            isnode_in_bound_nodes(nindex_to_node()->at(m))) {
+            is_node_in_configured_nodes(nindex_to_node()->at(m)) &&
+            is_node_in_bound_nodes(nindex_to_node()->at(m))) {
           distance = numa_distance(nindex_to_node()->at(i), nindex_to_node()->at(m));
           // If a closest node is found, update. There is always at least one
           // configured and bound node in the system so there is always at least
@@ -3030,9 +3035,13 @@
 os::Linux::numa_bitmask_isbitset_func_t os::Linux::_numa_bitmask_isbitset;
 os::Linux::numa_distance_func_t os::Linux::_numa_distance;
 os::Linux::numa_get_membind_func_t os::Linux::_numa_get_membind;
+os::Linux::numa_get_interleave_mask_func_t os::Linux::_numa_get_interleave_mask;
+os::Linux::NumaAllocationPolicy os::Linux::_current_numa_policy;
 unsigned long* os::Linux::_numa_all_nodes;
 struct bitmask* os::Linux::_numa_all_nodes_ptr;
 struct bitmask* os::Linux::_numa_nodes_ptr;
+struct bitmask* os::Linux::_numa_interleave_bitmask;
+struct bitmask* os::Linux::_numa_membind_bitmask;
 
 bool os::pd_uncommit_memory(char* addr, size_t size) {
   uintptr_t res = (uintptr_t) ::mmap(addr, size, PROT_NONE,
@@ -4944,6 +4953,74 @@
   OSContainer::init();
 }
 
+void os::Linux::numa_init() {
+
+  // Java can be invoked as
+  // 1. Without numactl and heap will be allocated/configured on all nodes as
+  //    per the system policy.
+  // 2. With numactl --interleave: 
+  //      Use numa_get_interleave_mask(v2) API to get nodes bitmask. The same
+  //      API for membind case bitmask is reset.
+  //      Interleave is only hint and Kernel can fallback to other nodes if 
+  //      no memory is available on the target nodes.
+  // 3. With numactl --membind: 
+  //      Use numa_get_membind(v2) API to get nodes bitmask. The same API for
+  //      interleave case returns bitmask of all nodes.
+  // numa_all_nodes_ptr holds bitmask of all nodes. 
+  // numa_get_interleave_mask(v2) and numa_get_membind(v2) APIs returns correct
+  // bitmask when externally configured to run on all or fewer nodes.
+
+  if (!Linux::libnuma_init()) {
+    UseNUMA = false;
+  } else {
+    if ((Linux::numa_max_node() < 1) || Linux::is_bound_to_single_node()) {
+      // If there's only one node (they start from 0) or if the process
+      // is bound explicitly to a single node using membind, disable NUMA.
+      UseNUMA = false;
+    } else {
+
+      LogTarget(Info,os) log;
+      LogStream ls(log);
+
+      Linux::set_configured_numa_policy(Linux::identify_numa_policy());
+
+      struct bitmask* bmp = Linux::_numa_membind_bitmask;
+      const char* numa_mode = "membind";
+
+      if (Linux::is_running_in_interleave_mode()) {
+        bmp = Linux::_numa_interleave_bitmask;
+        numa_mode = "interleave";
+      }
+
+      ls.print("UseNUMA is enabled and invoked in '%s' mode."
+               "Heap will be configured using NUMA memory nodes:", numa_mode);
+
+      for (int node = 0; node <= Linux::numa_max_node(); node++) {
+        if (Linux::_numa_bitmask_isbitset(bmp, node)) {
+          ls.print(" %d", node);
+        }
+      }
+    }
+  }
+
+  if (UseParallelGC && UseNUMA && UseLargePages && !can_commit_large_page_memory()) {
+    // With SHM and HugeTLBFS large pages we cannot uncommit a page, so there's no way
+    // we can make the adaptive lgrp chunk resizing work. If the user specified both
+    // UseNUMA and UseLargePages (or UseSHM/UseHugeTLBFS) on the command line - warn
+    // and disable adaptive resizing.
+    if (UseAdaptiveSizePolicy || UseAdaptiveNUMAChunkSizing) {
+      warning("UseNUMA is not fully compatible with SHM/HugeTLBFS large pages, "
+              "disabling adaptive resizing (-XX:-UseAdaptiveSizePolicy -XX:-UseAdaptiveNUMAChunkSizing)");
+      UseAdaptiveSizePolicy = false;
+      UseAdaptiveNUMAChunkSizing = false;
+    }
+  }
+
+  if (!UseNUMA && ForceNUMA) {
+    UseNUMA = true;
+  }
+}
+
 // this is called _after_ the global arguments have been parsed
 jint os::init_2(void) {
 
@@ -4988,32 +5065,7 @@
                Linux::glibc_version(), Linux::libpthread_version());
 
   if (UseNUMA) {
-    if (!Linux::libnuma_init()) {
-      UseNUMA = false;
-    } else {
-      if ((Linux::numa_max_node() < 1) || Linux::isbound_to_single_node()) {
-        // If there's only one node (they start from 0) or if the process
-        // is bound explicitly to a single node using membind, disable NUMA.
-        UseNUMA = false;
-      }
-    }
-
-    if (UseParallelGC && UseNUMA && UseLargePages && !can_commit_large_page_memory()) {
-      // With SHM and HugeTLBFS large pages we cannot uncommit a page, so there's no way
-      // we can make the adaptive lgrp chunk resizing work. If the user specified both
-      // UseNUMA and UseLargePages (or UseSHM/UseHugeTLBFS) on the command line - warn
-      // and disable adaptive resizing.
-      if (UseAdaptiveSizePolicy || UseAdaptiveNUMAChunkSizing) {
-        warning("UseNUMA is not fully compatible with SHM/HugeTLBFS large pages, "
-                "disabling adaptive resizing (-XX:-UseAdaptiveSizePolicy -XX:-UseAdaptiveNUMAChunkSizing)");
-        UseAdaptiveSizePolicy = false;
-        UseAdaptiveNUMAChunkSizing = false;
-      }
-    }
-
-    if (!UseNUMA && ForceNUMA) {
-      UseNUMA = true;
-    }
+    Linux::numa_init();
   }
 
   if (MaxFDLimit) {
diff -r 5d7e4d832868 src/hotspot/os/linux/os_linux.hpp
--- a/src/hotspot/os/linux/os_linux.hpp	Sat Jan 12 13:33:18 2019 +0100
+++ b/src/hotspot/os/linux/os_linux.hpp	Sat Jan 12 23:11:54 2019 +0530
@@ -211,6 +211,7 @@
   // none present
 
  private:
+  static void numa_init();
   static void expand_stack_to(address bottom);
 
   typedef int (*sched_getcpu_func_t)(void);
@@ -222,6 +223,7 @@
   typedef void (*numa_interleave_memory_func_t)(void *start, size_t size, unsigned long *nodemask);
   typedef void (*numa_interleave_memory_v2_func_t)(void *start, size_t size, struct bitmask* mask);
   typedef struct bitmask* (*numa_get_membind_func_t)(void);
+  typedef struct bitmask* (*numa_get_interleave_mask_func_t)(void);
 
   typedef void (*numa_set_bind_policy_func_t)(int policy);
   typedef int (*numa_bitmask_isbitset_func_t)(struct bitmask *bmp, unsigned int n);
@@ -239,9 +241,12 @@
   static numa_bitmask_isbitset_func_t _numa_bitmask_isbitset;
   static numa_distance_func_t _numa_distance;
   static numa_get_membind_func_t _numa_get_membind;
+  static numa_get_interleave_mask_func_t _numa_get_interleave_mask;
   static unsigned long* _numa_all_nodes;
   static struct bitmask* _numa_all_nodes_ptr;
   static struct bitmask* _numa_nodes_ptr;
+  static struct bitmask* _numa_interleave_bitmask;
+  static struct bitmask* _numa_membind_bitmask;
 
   static void set_sched_getcpu(sched_getcpu_func_t func) { _sched_getcpu = func; }
   static void set_numa_node_to_cpus(numa_node_to_cpus_func_t func) { _numa_node_to_cpus = func; }
@@ -255,10 +260,21 @@
   static void set_numa_bitmask_isbitset(numa_bitmask_isbitset_func_t func) { _numa_bitmask_isbitset = func; }
   static void set_numa_distance(numa_distance_func_t func) { _numa_distance = func; }
   static void set_numa_get_membind(numa_get_membind_func_t func) { _numa_get_membind = func; }
+  static void set_numa_get_interleave_mask(numa_get_interleave_mask_func_t func) { _numa_get_interleave_mask = func; }
   static void set_numa_all_nodes(unsigned long* ptr) { _numa_all_nodes = ptr; }
   static void set_numa_all_nodes_ptr(struct bitmask **ptr) { _numa_all_nodes_ptr = (ptr == NULL ? NULL : *ptr); }
   static void set_numa_nodes_ptr(struct bitmask **ptr) { _numa_nodes_ptr = (ptr == NULL ? NULL : *ptr); }
+  static void set_numa_interleave_bitmask(struct bitmask* ptr)     { _numa_interleave_bitmask = ptr ;   }
+  static void set_numa_membind_bitmask(struct bitmask* ptr)        { _numa_membind_bitmask = ptr ;      }
   static int sched_getcpu_syscall(void);
+
+  enum NumaAllocationPolicy {
+    NotInitialized,
+    Membind,
+    Interleave
+  };
+  static NumaAllocationPolicy _current_numa_policy;
+
  public:
   static int sched_getcpu()  { return _sched_getcpu != NULL ? _sched_getcpu() : -1; }
   static int numa_node_to_cpus(int node, unsigned long *buffer, int bufferlen) {
@@ -272,11 +288,33 @@
   static int numa_tonode_memory(void *start, size_t size, int node) {
     return _numa_tonode_memory != NULL ? _numa_tonode_memory(start, size, node) : -1;
   }
+
+  static bool is_running_in_interleave_mode() { 
+    return _current_numa_policy == Interleave; 
+  }
+
+  static void set_configured_numa_policy(NumaAllocationPolicy numa_policy) { 
+    _current_numa_policy = numa_policy ; 
+  }
+
+  static NumaAllocationPolicy identify_numa_policy() { 
+    for (int node = 0; node <= Linux::numa_max_node(); node++) {
+      if (Linux::_numa_bitmask_isbitset(Linux::_numa_interleave_bitmask, node)) {
+        return Interleave;
+      }
+    }
+    return Membind; 
+  }
+
   static void numa_interleave_memory(void *start, size_t size) {
-    // Use v2 api if available
-    if (_numa_interleave_memory_v2 != NULL && _numa_all_nodes_ptr != NULL) {
-      _numa_interleave_memory_v2(start, size, _numa_all_nodes_ptr);
-    } else if (_numa_interleave_memory != NULL && _numa_all_nodes != NULL) {
+    // Prefer v2 API
+    if (_numa_interleave_memory_v2 != NULL) {
+      if (is_running_in_interleave_mode()) {
+        _numa_interleave_memory_v2(start, size, _numa_interleave_bitmask);
+      } else if (_numa_membind_bitmask != NULL) {
+        _numa_interleave_memory_v2(start, size, _numa_membind_bitmask);
+      }
+    } else if (_numa_interleave_memory != NULL) {
       _numa_interleave_memory(start, size, _numa_all_nodes);
     }
   }
@@ -291,14 +329,14 @@
   static int get_node_by_cpu(int cpu_id);
   static int get_existing_num_nodes();
   // Check if numa node is configured (non-zero memory node).
-  static bool isnode_in_configured_nodes(unsigned int n) {
+  static bool is_node_in_configured_nodes(unsigned int n) {
     if (_numa_bitmask_isbitset != NULL && _numa_all_nodes_ptr != NULL) {
       return _numa_bitmask_isbitset(_numa_all_nodes_ptr, n);
     } else
       return false;
   }
   // Check if numa node exists in the system (including zero memory nodes).
-  static bool isnode_in_existing_nodes(unsigned int n) {
+  static bool is_node_in_existing_nodes(unsigned int n) {
     if (_numa_bitmask_isbitset != NULL && _numa_nodes_ptr != NULL) {
       return _numa_bitmask_isbitset(_numa_nodes_ptr, n);
     } else if (_numa_bitmask_isbitset != NULL && _numa_all_nodes_ptr != NULL) {
@@ -317,16 +355,19 @@
       return false;
   }
   // Check if node is in bound node set.
-  static bool isnode_in_bound_nodes(int node) {
-    if (_numa_get_membind != NULL && _numa_bitmask_isbitset != NULL) {
-      return _numa_bitmask_isbitset(_numa_get_membind(), node);
-    } else {
-      return false;
+  static bool is_node_in_bound_nodes(int node) {
+    if (_numa_bitmask_isbitset != NULL) {
+      if (is_running_in_interleave_mode()) {
+        return _numa_bitmask_isbitset(_numa_interleave_bitmask, node); 
+      } else {
+        return _numa_membind_bitmask != NULL ? _numa_bitmask_isbitset(_numa_membind_bitmask, node) : false;
+      }
     }
+    return false;
   }
   // Check if bound to only one numa node.
   // Returns true if bound to a single numa node, otherwise returns false.
-  static bool isbound_to_single_node() {
+  static bool is_bound_to_single_node() {
     int nodes = 0;
     struct bitmask* bmp = NULL;
     unsigned int node = 0;


More information about the hotspot-dev mailing list