Fix mis-memory counting in memory monitor for contaienr environment (#8113)

Co-authored-by: weich <weich@nvidia.com>
2026-06-30 05:41:19 +08:00 · 2020-04-22 15:32:35 -06:00
parent 0bb918f2b1
commit 51559c08b9
1 changed files with 7 additions and 0 deletions
@@ -114,6 +114,13 @@ class MemoryMonitor:
                with open("/sys/fs/cgroup/memory/memory.usage_in_bytes",
                          "rb") as f:
                    used_gb = int(f.read()) / (1024**3)
+                # Exclude the page cache
+                with open("/sys/fs/cgroup/memory/memory.stat", "r") as f:
+                    for line in f.readlines():
+                        if line.split(" ")[0] == "cache":
+                            used_gb = \
+                                used_gb - int(line.split(" ")[1]) / (1024**3)
+                assert used_gb >= 0
            if used_gb > total_gb * self.error_threshold:
                raise RayOutOfMemoryError(
                    RayOutOfMemoryError.get_message(used_gb, total_gb,