RFE (m) (Prelminary): JDK-7197666: java -d64 -version core dumps in a box with lots of memory

[Bengt Rutisson]

Hi Coleen,

On 4/4/13 5:25 PM, Coleen Phillimore wrote:
>
> On 04/04/2013 07:32 AM, Bengt Rutisson wrote:
>>
>>
>> Hi Colleen,
>>
>> Thank you so much for looking at this!
>>
>> On 4/1/13 11:04 PM, Coleen Phillimore wrote:
>>>
>>> Bengt,
>>>
>>> This is a nice way to solve this, although don't know why you don't 
>>> just use mmap for all the task queues.   You don't allocate and 
>>> deallocate them a lot, do you?
>>
>> Right, I was thinking about using mmap directly but when I looked at 
>> the current usages of ::mmap I can only find that it is being used in 
>> src/os files. So, it seems to me like this is not the API we are 
>> supposed to use in the rest of the code. The next level of 
>> abstraction is os::reserve_memory() and os::commit_memory(), so 
>> that's why I picked them. It would have been nice if I could have 
>> just used os::commit_memory(), but internally it uses MAP_FIXED which 
>> means that I can't pass NULL as the addr parameter. So, I have to use 
>> os::reserve_memory() first.
>>
>> Are you ok with these calls, or do you think I should change to use 
>> mmap directly?
>
> Oh, no, I think you should use os::reserve_memory().   My question was 
> why do you conditionally use reserve_memory vs. os::malloc() for these 
> task queues.  Or are they always big enough that they'll end up using 
> os::reserve_memory().

Ah. Sorry, I misunderstood before.

Good point. I think the task queues will always be large but I would 
like the ArrayAllocator to be of more general use. We could use it for 
structures that have a dynamic size and only need to be mmapped if they 
are large enough.

>>
>>>
>>> You can also use the trick in globals.hpp:
>>>
>>> *+   product(uintx, ArrayAllocatorMallocLimit, SOLARIS_ONLY(64K) NOT_SOLARIS(MAX_UINT),  \*
>>> *+           "Allocation less than this value will be allocated "              \*
>>> *+           "using malloc. Larger allocations will use mmap.")*
>>>
>>> And include both solaris sparc and x86 in the size calculation 
>>> (since sparc malloc has the same problem as x86 just not as bad), 
>>> and save editing all the globals_<os_cpu>.hpp files.
>>
>> Thanks! This is much better! It also means that I can use an 
>> experimental flag, so I'll change to that.
>
> Yes, I guess it's better as an experimental flag.  It could be product 
> also.  People might want their own sizes.

I'd like to keep it experimental for a while to avoid that it is being 
used too much.

Thanks,
Bengt

>
> Coleen
>
>>
>>
>> Thanks again for this feedback. It was exactly what I needed. I'll 
>> send out an updated webrev and a "real" review request soon.
>>
>> Bengt
>>
>>>
>>> Coleen
>>>
>>> On 03/28/2013 06:09 PM, Bengt Rutisson wrote:
>>>>
>>>> Hi all,
>>>>
>>>> Sending this to both runtime and GC since I think it concerns both 
>>>> areas.
>>>>
>>>> I'd like some feedback on this preliminary change. I still want to 
>>>> do some more testing and evaluation before I ask for final reviews:
>>>>
>>>> http://cr.openjdk.java.net/~brutisso/7197666/webrev.00/
>>>>
>>>> In particular I would like some feedback on these questions:
>>>>
>>>> - I am adding a flag that has the same value on all platforms 
>>>> except Solaris x86. There is the product_pd flag macro to support 
>>>> this. But there is no experimental_pd marcro. I would have 
>>>> preferred to make my new flag experimental. Should I add 
>>>> experimental_pd or should I just use a product flag?
>>>>
>>>> - Even with product_pd I think I still have to go in to all the 
>>>> different platform files and add the exact same code to give the 
>>>> flag a default value on all platforms. Is there a way to have a 
>>>> default value and only override it on Solaris x86?
>>>>
>>>> - The class I am adding, ArrayAllocator, wants to choose between 
>>>> doing malloc and mmap. Normally we use ReservedSpace and 
>>>> VirtualSpace to get mapped memory. However, those classes are kind 
>>>> of clumsy when I just want to allocate one chunk of memory. It is 
>>>> much simpler to use the os::reserve_memory() and 
>>>> os::commit_memory() methods directly. I think my use case here 
>>>> motivate using these methods directly, but is there some reason not 
>>>> to do that?
>>>>
>>>> Some background on the change:
>>>>
>>>> The default implementation of malloc on Solaris has several 
>>>> limitation compared to malloc on other platforms. One limitation is 
>>>> that it can only use one consecutive chunk of memory. Another 
>>>> limitation is that it always allocates in this single chunk of 
>>>> memory no matter how large the requested amount of memory is. Other 
>>>> malloc implementations normally use mapped memory for large 
>>>> allocations.
>>>>
>>>> The Java heap is mapped in memory and we try to pick a good address 
>>>> for it. The lowest allowed address is controlled by 
>>>> HeapBaseMinAddress. This is only 256 MB on Solaris x86 (other 
>>>> platforms have at least 2 GB). Since the C heap ends up below the 
>>>> Java heap it means that in some cases it is limited to 256 MB.
>>>>
>>>> When we run with ParallelOldGC we get three task queues per GC 
>>>> thread. Each task queue takes mallocs 1MB. The failing machine in 
>>>> the bug report has lots of CPUs and ends up with 83 GC threads. 
>>>> This is 249 MB, which is more than we can get out of the 256 MB 
>>>> limited C heap considering that there are other things that get 
>>>> malloced too.
>>>>
>>>> So, the problems occur mostly on Solaris x86. My suggested fix 
>>>> tries to address this by letting the task queues be mapped instead 
>>>> of malloced on Solaris x86. Instead of inlining this logic in 
>>>> taskqueue.cpp I added a more general class. The reason for this is 
>>>> that I think we need to use the same logic in more places, 
>>>> especially for G1, which is mallocing quite a lot.
>>>>
>>>> Since I think malloc on other platforms use mapped memory for large 
>>>> malloc requests I think it is enough for this change to have effect 
>>>> on Solaris. The other platforms probably have better heuristics 
>>>> than I can come up with for which sizes should be mapped. On Sparc 
>>>> we have the same limitation with malloc, but we have more memory 
>>>> available for the C heap. This is why I have only enabled this for 
>>>> Solaris x86.
>>>>
>>>> Also, I will be on vacation for a few days. Back in the office 
>>>> Thrusday April 4. I'm happy for any feedback on this, but if I 
>>>> don't respond until next week you know why :)
>>>>
>>>> Thanks,
>>>> Bengt
>>>
>>
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/hotspot-runtime-dev/attachments/20130405/bf318494/attachment.html