[12] RFR(S): 8177899: Tests fail due to code cache exhaustion on machines with many cores

[Vladimir Kozlov]

Hi Tobias,

On 10/24/18 3:37 AM, Tobias Hartmann wrote:
> Hi Vladimir,
> 
> thanks for looking at this!
> 
> On 23.10.18 18:46, Vladimir Kozlov wrote:
>> "The easy fix would be to add -XX:CICompilerCount=2 to the affected tests. but we might also want to
>> think about changing the default value of CICompilerCount to take code cache size into consideration
>> and not just number of cpus."
>>
>> Even with UseDynamicNumberOfCompilerThreads it is possible to use all CICompilerCount compiler
>> threads when there are a lot of request. And code respect this number.
> 
> Yes we can do that but it might still happen that we run out of space (see below). Here's a new
> webrev that lowers the CICompilerCount if we don't have enough space in the code cache:
> http://cr.openjdk.java.net/~thartmann/8177899/webrev.01/

Good. I think you need to check if CICompilerCount is FLAG_IS_DEFAULT when changing it.

> 
>> I think scaling down CICompilerCount based on CodeCache size is preferable. Unless CICompilerCount
>> is specified on command line - in which case we should issue a error as we do now but more verbose
>> including CICompilerCount.
> 
> There are several places where code cache allocation may still fail at runtime (for example, if
> classes are loaded and adapters are created). I don't think we should be more verbose there than the
> usual out of code cache space message. There are other flags that can be set by the user and affect
> code cache consumption as well (for example, compilation thresholds and sweeper settings). So we
> would need to include all of them. Sometimes we don't even print to the command line but throw a
> java.lang.VirtualMachineError ("Out of space in CodeCache for method handle intrinsic").

Okay.

> 
> The check in codeCache.cpp is still too strong. Compiler thread buffers can also go into a nmethod
> code heap if the non-nmethod code heap is full. We should only check if we have at least

Yes, it is true. I forgot about that.

> CodeCacheMinimumUseSpace (otherwise we might crash at startup).

Thanks,
Vladimir

> 
> Thanks,
> Tobias
>