The SPECjbb2005 issue is a race condition in the classpath xml implementation. It was only being exposed on our new dual core machines. Our longer standing suspicions of rvmx86lnx64 not acting like a dual processor may be why this issue hasn’t shown up on rvmx86lnx64.

 

This has been committed as a patch against classpath in revision 12395 and is tracked in issue 1728742 while we go through the process of getting classpath fixed.

 

Cheers,

Daniel.

 


From: jikesrvm-core-bounces@lists.sourceforge.net [mailto:jikesrvm-core-bounces@lists.sourceforge.net] On Behalf Of Steve Blackburn
Sent: Thursday, 31 May 2007 10:50 AM
To: Automatic mail messages and subsequent discussion of nightly regression runs
Cc: jikesrvm-core@lists.sourceforge.net
Subject: Re: [Jikesrvm-core] [Jikesrvm-regression] FAILED 203 tests

 

 

On 31/05/2007, at 12:23 AM, David P Grove wrote:

       I'm not seeing that.  I think you are perhaps confusing results from x86lnx32 and x86lnx64.   

 

You're right.  My mistake.

 

It turns out that that is the first full sanity run from rvmx86lnx32 since it was upgraded.  Two things have changed: a) the OS and libraries have been upgraded to Ubuntu's Feisty Fawn, and b) the hardware has been upgraded to a Core 2 Duo.

 

Most of the failures are due to problems with the newer version of xvfb (virtual frame buffer), which both chart and jbb's reporters use.  Robin has a good work-around for that which should be in place RSN.

 

However, Daniel has also noticed a real regression on jbb in that new environment (on a different machine with the same configuration). :-/  It is not clear whether jbb has actually regressed or whether it is just that that new environment exposes a bug we did not see before.  Aside from the possibility of issues with libraries (which is real since the problems appear to relate to xml handling), the other thing is that this new machine is the fastest CMP we've tested Jikes RVM on so far, so perhaps race conditions will be more exposed.

 

--Steve