Michael Nelson created CASSANDRA-7931:
-----------------------------------------

             Summary: Single Host CPU profiling data dump
                 Key: CASSANDRA-7931
                 URL: https://issues.apache.org/jira/browse/CASSANDRA-7931
             Project: Cassandra
          Issue Type: Improvement
         Environment: 2.1.0 on my local machine
            Reporter: Michael Nelson
            Priority: Minor
         Attachments: traces.txt

At Boot Camp today I did some CPU profiling and wanted to turn in my findings 
somehow. So here they are in a JIRA.

I ran cassandra-stress read test against my local machine (single node: SSD, 8 
proc, plenty of RAM). I'm using straight 2.1.0 with the "Lightweight Java 
Profiler" by Jeremy Manson 
(https://code.google.com/p/lightweight-java-profiler/). This is a lower-level 
profiler that more accurately reports what the CPU is actually doing rather 
than where time is going.

Attached is the report. Here are a few high points with some commentary from 
some investigation I did today:

1. The number one _consumer_of_CPU_ is this stack trace:
sun.misc.Unsafe.park(Unsafe.java:-1)
java.util.concurrent.locks.LockSupport.parkNanos(LockSupport.java:349)
org.apache.cassandra.concurrent.SEPWorker.doWaitSpin(SEPWorker.java:236)
org.apache.cassandra.concurrent.SEPWorker.run(SEPWorker.java:74)
java.lang.Thread.run(Thread.java:745)

There is a CPU cost to going into and out of Unsafe.park() and the SEPWorkers 
are doing it so much when they are spinning that they were the largest single 
consumer of CPU. Note that this profiler is sampling the CPU so it isn't the 
_blocking_ that this call does that is the issue. It is the CPU actually being 
consumed as it prepares to block and then later comes out. This kind of lower 
level fine grained control of threads is tricky and there appears to be room 
for improvement here. I tried a few approaches that didn't pan out. There 
really needs to be a way for threads to block waiting for the executor queues. 
That way they will be immediately available (rather than waiting for them to 
poll the executors) and will not consume CPU when they aren't needed. Maybe 
block for some short period of time and then become available for other queues 
after that?

2. Second is Netty writing to sockets. I didn't investigate this. Netty is 
pretty optimized. Someone mentioned there are native options for Netty. 
Probably worth trying.

3. Third is similar to #1:
sun.misc.Unsafe.park(Unsafe.java:-1)
java.util.concurrent.locks.LockSupport.parkNanos(LockSupport.java:349)
org.apache.cassandra.utils.concurrent.WaitQueue$AbstractSignal.awaitUntil(WaitQueue.java:303)
     
org.apache.cassandra.utils.concurrent.SimpleCondition.await(SimpleCondition.java:63)
org.apache.cassandra.service.ReadCallback.await(ReadCallback.java:90)
org.apache.cassandra.service.ReadCallback.get(ReadCallback.java:100)
org.apache.cassandra.service.AbstractReadExecutor.get(AbstractReadExecutor.java:144)
org.apache.cassandra.service.StorageProxy.fetchRows(StorageProxy.java:1226)

I looked at this a little. The asynchronous callback using Unsafe.park() is 
being invoked even for local reads (like my local laptop). It only rarely 
blocked. It just paid the CPU cost of going in and immediately coming out 
because the read was already satisfied from the buffer cache. :-(

4. Fourth is Netty doing epoll. Again, worth considering Netty's native 
optimizations to help here.

The others are interesting but, for this little benchmark, not as big a deal. 
For example, ArrayList.grow() shows up more than I would have expected.

The Boot Camp was great. Feel free to contact me if I can help somehow.




--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to