SapMachine (21) #2109: Linux: Configuration of glibc malloc arenas should be possible

Author: TheRealMDoerr

src/hotspot/os/linux/globals_linux.hpp

@@ -112,6 +112,11 @@
          "omitted.\n"                                                  \
          "Example: \"-XX:HiMemReportExec=GC.class_histogram -all;GC.heap_dump\"") \
                                                                         \
+  /* SapMachine 2025-11-24: Configurable limit of malloc arenas */      \
+  product(int, GlibcMallocArenas, 0,                                    \
+          "Limit glibc malloc arenas, 0 means use OS default, "         \
+          "1 minimizes memory utilization (our default in later releases)") \
+                                                                        \
   product(bool, UseCpuAllocPath, false, DIAGNOSTIC,                     \
           "Use CPU_ALLOC code path in os::active_processor_count ")     \
                                                                         \

src/hotspot/os/linux/os_linux.cpp

@@ -4689,6 +4689,21 @@ jint os::init_2(void) {
 
   Linux::fast_thread_clock_init();
 
+  // SapMachine 2025-11-24:
+  // By default, glibc allocates a new 64 (or 128) MB malloc arena for every
+  // thread (up to a certain limit which is typically 8 * processor count).
+  // This is good for few threads which perform a lot of concurrent mallocs,
+  // but it doesn't fit well to the JVM which has its own memory management
+  // and rather allocates fewer and larger chunks.
+  // Using only one arena significantly reduces virtual memory footprint and
+  // fragmentation. Saving memory seems to be more valuable for the JVM than
+  // optimizing concurrent mallocs.
+#ifdef __GLIBC__
+  if (GlibcMallocArenas > 0) {
+    mallopt(M_ARENA_MAX, GlibcMallocArenas);
+  }
+#endif
+
   if (PosixSignals::init() == JNI_ERR) {
     return JNI_ERR;
   }