Garbage Collection Pauses & Non-interruptable System Calls

Tue Apr 14 18:33:48 UTC 2009

After reading the last paragraph of HotSpot Runtime Overview of JNI more 
closely, I understand more.  I think we're almost on the same page.  The 
problem seems to be that all threads are suspended until the Java thread 
returns from the native call.

Mark Maxey
Raytheon, Garland
580/2/P22-1
(972)205-5760
Mark_R_Maxey at Raytheon.com

Mark R Maxey/US/Raytheon
04/14/2009 12:37 PM

To
"Y. Srinivas Ramakrishna" <Y.S.Ramakrishna at Sun.COM>
cc
hotspot-gc-dev at openjdk.java.net, Y.S.Ramakrishna at Sun.COM, Andrew M 
Dungan/US/Raytheon at MAIL, David A Lilly/RCS/Raytheon/US at MAIL, Mark R 
Maxey/US/Raytheon at MAIL
Subject
Re: Garbage Collection Pauses & Non-interruptable System Calls

Thank you for your reply.  The speed and depth of your response is 
encouraging.

Let me confess something I should have done up-front.  The behavior we're 
seeing is using JDK 5 via JRockit R27.6.  We're in the process of 
reproducing these problems under HotSpot JDK 6 Update 12, though it'll be 
a few days before we can do so.  The reason I'm pinging this forum is to 
research in advance what differences we might expect between the two JVMs.

Let me describe exactly what we're seeing as provided by doing an strace 
on the process:
A Java thread calls a native C code that ultimately calls a pwrite().  We 
suspect that the device driver ultimately makes a non-interruptable system 
call to transfer the data directly from our mem-aligned 128 MB buffer to 
disk.
The GC thread sends a tgkill(SIGUSR1) to all threads
The GC thread waits on mutex #1 (presumably waiting on all the threads to 
signal it that it can begin GC)
The Java thread wakes mutex #1 (presumably signaling the GC it is ready to 
go)
The Java thread waits on mutex #2 (presumably waiting on GC to finish)
The GC thread wakes mutex #2 (presumably telling the Java thread it can 
resume processing)

We're seeing times between #3 & #4 that are proportional to the amount of 
time spent in the pwrite().  We also see some overhead between #5 &#6 that 
is proportional to the number of Java threads we have (currently between 
30 & 40 that we've created not counting the JVMs).

Unfortunately, the JRockit logging only reveals the actual time GC takes 
(#4 - #5).  Hopefully, HotSpot's logging includes the total time (#2 - 
#6).

I'm pursuing these questions with Oracle/BEA.  Again, I'm just trying get 
a feel for HotSpot's behavior in comparison.  While we're using JRockit 
today, HotSpot will be our ultimate platform.

One alternate solution that has been suggested is infrequently calling GC 
explicitly within our code during special times when we know we can afford 
to take the hit.  We would even accept a greater hit than normal if we 
could avoid being impacted during critical times.   Everything I've ever 
read says to not do this, but I'm curious why in this case this is a bad 
idea.  Note that we're using the concurrent GC, so I'm not even sure if 
System.gc() supports this.

Thanks again!

Mark Maxey
Raytheon, Garland
580/2/P22-1
(972)205-5760
Mark_R_Maxey at Raytheon.com

"Y. Srinivas Ramakrishna" <Y.S.Ramakrishna at Sun.COM> 
Sent by: Y.S.Ramakrishna at Sun.COM
04/14/2009 10:19 AM

To
Mark R Maxey <Mark_R_Maxey at raytheon.com>
cc
hotspot-gc-dev at openjdk.java.net
Subject
Re: Garbage Collection Pauses & Non-interruptable System Calls

Hello Mark --

I am assuming your threads doing DMA are actually executing native code 
(or
waiting for signals in native code).  Threads in native code do not need 
to
synchronize \in any manner with GC while they are executing native code.
It is only the transitions to and from native mode (from Java code) that 
require
synchronization. Roughly speaking, the JVM fences off those native
threads so that, in the event that they need to re-enter the JVM or
access the Java heap, they will be suspended until a GC/safepoint that
is in progress is completed.

Thus, I do not believe you need to fear that a long-running DMA call would
cause GC's to be delayed (which I understand is your  main concern below).

Have you actually seen cases where this is happening? If so, what 
version of the JDK
are you running?

thanks.
-- ramki

Mark R Maxey wrote:
> Hello,
>
> I have a problem I was hoping with which I need some advice.
>
> We wrote a custom JNI library for file I/O that sits underneath the Java 

> NIO FileChannel.  One of our driving requirements is highly performant 
> file I/O.  We achieved this by doing DMA I/O from large direct memory 
> aligned buffers.  The JNI is very trivial - it just takes a buffer and 
> performs the appropriate system call based on the parameters given to 
it. 
> 100% of the logic for calculating offsets, buffer management, etc. is 
all 
> in our implementation of java.nio.FileChannel.
>
> Here's our problem:  We have requirements to respond to some messages in 

> as little as 250 ms.  During this time, we're doing file writes of 128 
MB 
> that take around 200 ms.  When GC kicks in, it tries to pause all 
threads. 
>  Because the DMA write is non-interruptable, GC waits for the I/O to 
> complete before being able to pause the thread & run.  That means that 
GC 
> can take well over 200 ms putting us in grave danger of missing our 
> timelines.  Worse, there is always the chance the write will hang due to 
a 
> bad filesystem.   We've seen this cause the JVM to hang indefinitely 
> forcing us to cycle the process.
>
> Unless we find a solution that allows GC to continue while doing this 
I/O, 
> we will convert all the code to C++.  While that might solve our 
timeline 
> for that particular process, we have many less performance critical 
> processes that use our JNI FileChannel libraries that would hang if a 
> filesystem goes bad.
>
> We've tweaked the file system device timeouts down to a minimum, but 
they 
> are still very high (on the order of several seconds to minutes).  It 
> would be nice if the JVM had a similar timeout for pausing threads, 
i.e., 
> where the pause times out after X number of milliseconds.  We'd be 
willing 
> to sacrifice a larger heap size and postpone GC in the hopes that the 
next 
> time it ran GC, we wouldn't be in the middle of a non-interruptable 
system 
> call.
>
> The only solution being batted around here is pushing the system calls 
out 
> of Java threads and into native threads.  The JNI call would push the 
info 
> for the I/O call onto a native C++ queue where a small number of native 
> threads (3?) would pull the data off the queue and perform the actual 
> system call.   The trick is finding an implementation where the Java 
> thread blocked waiting on a response from the native thread is 
> interruptible.  All this assumes GC doesn't try to pause native threads. 

> We thought about using pthreads, but were concerned about its signal 
> interaction with the JVM.  So, we're leaning towards using pipes to push 

> data from one thread to another.
>
> If you have any suggestions or advice, we are desperate for your wisdom.
>
> Thanks!
>
>
> Mark Maxey
> Raytheon, Garland
> 580/2/P22-1
> (972)205-5760
> Mark_R_Maxey at Raytheon.com
>
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/hotspot-gc-dev/attachments/20090414/5c98129f/attachment.htm>