RFR: 8195621: JTReg failure in gc/g1/TestLargePageUseForAuxMemory.java

[stewartd.qdt]

Looking at the test, I think we can remove the -Xms flags from the test. In doing this, the test passes and it still seems to generate the right page sizes, which appears to be what the test was looking for. 

I have updated the webrev:
http://cr.openjdk.java.net/~dstewart/8195621/webrev.01/

Please let me know if this is acceptable. 

Daniel

-----Original Message-----
From: Thomas Schatzl [mailto:thomas.schatzl at oracle.com] 
Sent: Wednesday, January 24, 2018 4:17 AM
To: stewartd.qdt <stewartd.qdt at qualcommdatacenter.com>; David Holmes <david.holmes at oracle.com>; hotspot-dev at openjdk.java.net
Subject: Re: RFR: 8195621: JTReg failure in gc/g1/TestLargePageUseForAuxMemory.java

Hi,

On Fri, 2018-01-19 at 15:11 +0000, stewartd.qdt wrote:
> Yes, I too wondered if the test should be changed to be a bit more 
> discerning in how it operates.
> 
> I agree it would be great if a member of the GC team could chime in on 
> this issue.
> 
> Daniel
> 
> -----Original Message-----
> 
> > This fix adds a function to WhiteBox to get the physical memory and 
> > then uses that to ensure we are not trying to allocate more than the 
> > physical amount of memory for our heap. Please let me know if this 
> > is a bad approach and I'll be happy to change as required.
> 
> This seems an improvement though I would expect the test to fail long 
> before we get close to the amount of physical memory available.
> Should the test be adjusting what it does based on the page size being 
> used, in conjunction with the amount of "physical memory"?
> 
> Hopefully one of the GC folk will chime in.

The test has obviously been written with existing test setups we have that are configured accordingly.

I think for the test it may be sufficient to simply remove the -Xms flags in the VM invocations. Then the test would not try to actually commit all of the pages for e.g. the heap.

That assumes it is no problem for these systems to only reserve 512G, and the systems themselves, if configured to use 1g huge pages is configured properly, i.e. there are actually have a few gb of memory available (unless -version does not even create objects on the java heap).

Otherwise, instead of adding a new whitebox method, it might be preferable to get free/physical memory using standard Java means like OperatingSystemMXBean [1].

Thanks,
  Thomas

[1] https://docs.oracle.com/javase/9/docs/api/com/sun/management/Operat
ingSystemMXBean.html