Details
-
Improvement
-
Status: Open
-
Major
-
Resolution: Unresolved
-
0.20.0
-
None
-
None
-
Linux, running Yahoo-based 0.20
-
Increases the accept() rate on the IPC socket, to avoid TCP connection refused errors on busy namenodes and jobtrackers that get a lot of connections.
Description
We were recently seeing issues in our environments where HDFS clients would experience RST's from the NN when trying to do RPC to get file info, which would cause the task to fatal out. After some debugging we identified this to be that the IPC server listen queue – ipc.server.listen.queue.size – was far too low, we had been using the default value of 128 and found we needed to bump it up to 10240 before resets went away (although this value is a bit suspect, as I will explain later in the issue).
When a large map job starts, lots of clients very quickly start to issue RPC requests to the namenode, which creates this listen queue filling up problem, because clients are opening connections faster than Hadoop's RPC server can process them. We went back to our 0.17 cluster and instrumented that with tcpdump and found that we had been sending RST's for a long time there, but the retry handling was implemented differently back in 0.17 so a single TCP failure wasn't task-fatal.
In our environment we have our TCP stack set to explicitly send resets when the listen queue gets overflowed (syctl net.ipv4.tcp_abort_on_overflow = 1), default linux behavior is to start dropping SYN packets and let the client retransmit. Other people may be experiencing this issue and not noticing it because they are using the default behavior, which is to let the NN drop packets on the floor and let clients retransmit.
So we've identified (at least) 3 improvements that can be made here:
1) In src/core/org/apache/hadoop/ipc/Server.java, Listener.doAccept() is currently hardcoded to do 10 accept()'s at a time, then it will start to read. We feel that it would be better to allow the server to be configured to support more than 10 accept's at one time using a configurable parameter. We can still leave 10 as the default.
2) Increase the default value of ipc.server.listen.queue.size from 128, or at least document that people with larger clusters starting thousands of mappers at once should increase this value. I wonder if a lot of people running larger clusters are dropping packets and don't realize it because TCP is covering them up. One one hand, yay TCP, on the other hand, those are needless delays and retries because the server can handle more connections.
3) Document that ipc.server.listen.queue.size may be limited to the value of SOMAXCONN (linux sysctl net.core.somaxconn ; default 4096 on our systems). The Java docs are not completely clear about this, and it's difficult to test because you can't query the backlog of a listening socket. We were under some time pressure in our case and tried 1024 which was not enough, and 10240 which worked, so we stuck with that.