[YARN-8202] DefaultAMSProcessor should properly check units of requested custom resource types against minimum/maximum allocation

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Fix Version/s: 3.1.1, 2.10.1
• Target Version/s: 3.1.1
• Hadoop Flags: Reviewed
• Assignee: Szilard Nemeth
• Reporter: Szilard Nemeth
• Created: 24/Apr/18 15:52
• Updated: 09/Dec/19 20:00
• Resolved: 10/May/18 16:52

Description

 

When I execute a pi job with arguments: 

-Dmapreduce.map.resource.memory-mb=200 -Dmapreduce.map.resource.resource1=500M 1 1000

and I have one node with 5GB of resource1, I get the following exception on every second and the job hangs:

2018-04-24 08:42:03,694 INFO org.apache.hadoop.ipc.Server: IPC Server handler 20 on 8030, call Call#386 Retry#0 org.apache.hadoop.yarn.api.ApplicationMasterProtocolPB.allocate from 172.31.119.172:58138

org.apache.hadoop.yarn.exceptions.InvalidResourceRequestException: Invalid resource request, requested resource type=[resource1] < 0 or greater than maximum allowed allocation. Requested resource=<memory:200, vCores:1, resource1: 500M>, maximum allowed allocation=<memory:6144, vCores:8, resource1: 5G>, please note that maximum allowed allocation is calculated by scheduler based on maximum resource of registered NodeManagers, which might be less than configured maximum allocation=<memory:8192, vCores:8192, resource1: 9223372036854775807G>
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.SchedulerUtils.validateResourceRequest(SchedulerUtils.java:286)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.SchedulerUtils.normalizeAndValidateRequest(SchedulerUtils.java:242)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.SchedulerUtils.normalizeAndvalidateRequest(SchedulerUtils.java:258)
        at org.apache.hadoop.yarn.server.resourcemanager.RMServerUtils.normalizeAndValidateRequests(RMServerUtils.java:249)
        at org.apache.hadoop.yarn.server.resourcemanager.DefaultAMSProcessor.allocate(DefaultAMSProcessor.java:230)
        at org.apache.hadoop.yarn.server.resourcemanager.scheduler.constraint.processor.DisabledPlacementProcessor.allocate(DisabledPlacementProcessor.java:75)
        at org.apache.hadoop.yarn.server.resourcemanager.AMSProcessingChain.allocate(AMSProcessingChain.java:92)
        at org.apache.hadoop.yarn.server.resourcemanager.ApplicationMasterService.allocate(ApplicationMasterService.java:433)
        at org.apache.hadoop.yarn.api.impl.pb.service.ApplicationMasterProtocolPBServiceImpl.allocate(ApplicationMasterProtocolPBServiceImpl.java:60)
        at org.apache.hadoop.yarn.proto.ApplicationMasterProtocol$ApplicationMasterProtocolService$2.callBlockingMethod(ApplicationMasterProtocol.java:99)
        at org.apache.hadoop.ipc.ProtobufRpcEngine$Server$ProtoBufRpcInvoker.call(ProtobufRpcEngine.java:523)
        at org.apache.hadoop.ipc.RPC$Server.call(RPC.java:991)
        at org.apache.hadoop.ipc.Server$RpcCall.run(Server.java:872)
        at org.apache.hadoop.ipc.Server$RpcCall.run(Server.java:818)
        at java.security.AccessController.doPrivileged(Native Method)
        at javax.security.auth.Subject.doAs(Subject.java:422)
        at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1682)
        at org.apache.hadoop.ipc.Server$Handler.run(Server.java:2678)

This is because org.apache.hadoop.yarn.server.resourcemanager.scheduler.SchedulerUtils#validateResourceRequest does not take resource units into account.

 

However, if I start a job with arguments: 

-Dmapreduce.map.resource.memory-mb=200 -Dmapreduce.map.resource.resource1=1G 1 1000

and I still have 5GB of resource1 on one node then the job runs successfully.

 

I also tried a third job run, when I request 1GB of resource1 and I have no nodes with any amount of resource1, then I restart the node with 5GBs of resource1, the job ultimately completes, but just after the node with enough resources registered in RM, which is the desired behaviour.

 

Issue Links

• is related to: YARN-7739 DefaultAMSProcessor should properly check customized resource types against minimum/maximum allocation