Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
-
None
Description
We have seen issues on our clusters where the current way of computing CPU usage is having float-arithmetic inaccuracies (the bug is still there in trunk)
Simple program to illustrate:
public static void main(String[] args) throws Exception { float result = 0.0f; for (int i = 0; i < 7; i++) { if (i == 6) { result += (float) 4 / (float)18; } else { result += (float) 2 / (float)18; } } for (int i = 0; i < 7; i++) { if (i == 6) { result -= (float) 4 / (float)18; } else { result -= (float) 2 / (float)18; } } System.out.println(result); }
// Printed
4.4703484E-8
2017-04-12 05:43:24,014 INFO org.apache.hadoop.yarn.server.nodemanager.containermanager.monitor.ContainersMonitorImpl: Not enough cpu for [container_e3295_1491978508342_0467_01_000030], Current CPU Allocation: [0.8888891], Requested CPU Allocation: [0.11111111]
There are a few places with this issue:
1. ResourceUtilization.java - set/getCPU both use float. When ContainerScheduler calls ContainersMonitor.increase/decreaseResourceUtilization, this may lead to issues.
2. AllocationBasedResourceUtilizationTracker.java - hasResourcesAvailable uses float as well for CPU computation.