CRR (S): 7099849: G1: include heap region information in hs_err files

Srinivas Ramakrishna ysr1729 at gmail.com
Wed Nov 2 16:38:39 UTC 2011


Ok to yr second and third points. The first may be kinda moot at this point :-)

-- ramki

Sent from my iPad

On Nov 2, 2011, at 7:44 AM, Tony Printezis <tony.printezis at oracle.com> wrote:

> Ramki,
> 
> I'd rather not. First, we'd need to introduce a new -XX parameter for this. :-) Second, customers should not have many hs_err files (does HotSpot crash this often?). And if they decide to turn the per-region output off, we (or whoever tries to track down what went wrong) would miss that helpful information. I'd rather always generate the per-region output.
> 
> Tony
> 
> On 11/1/2011 5:07 PM, Srinivas Ramakrishna wrote:
>> Tony, What if you gave users the choice.... If they feel the g1 hserr files are too big for comfort they could explicitly switch the extended printing off, perhaps by explicitly setting An appropriate parameter off... Or may be deal with that issue if we get complaints to that effect from users...
>> 
>> -- ramki
>> 
>> Sent from my iPad
>> 
>> On Nov 1, 2011, at 6:23 AM, Bengt Rutisson<bengt.rutisson at oracle.com>  wrote:
>> 
>>> Tony,
>>> 
>>> On 2011-10-31 18:24, Tony Printezis wrote:
>>>> ....
>>>> On 10/31/2011 5:55 AM, Bengt Rutisson wrote:
>> ....
>>>>> 
>>>>> 
>>>>> 
>>>>> 
>>>>> 
>>>>> 
>>>>> 
>>>> 
>>>> 
>>>> 
>>>> 
>>>> 
>>>> 
>>>> 
>>>>> Finally, just for the record, we have tests that run without setting -Xms or -Xmx. If such a test ends up on a large machine we can end up with many more than 1000-2000 regions. I have seen ten times as many regions.
>>>> I could add a max of how many per-region entries we'll write to the hs_err file and basically skip the rest (by adding an appropriate warning in the log of course). Having said that, this might make this output useless given that the entries we'll skip might be the ones that we'd need. So, I'm inclined not to do that.
>>> I agree. A maximum will likely reduce the benefit of this fix.
>>> 
>>>>> It might be unlikely that customers run without limiting the heap size on this large machines, but we will run into hs_err files with 10s of thousands of heap regions in our own testing. Personally I think the information is useful enough for this to be acceptable.
>>>> Agreed.
>>>> 
>>>>> One alternative would of course be to log this in a separate hs_err_g1 file...
>>>> I really don't think we need to go there! We'll generate the same amount of output but now split over two files! And the customers will have to remember to also keep around and send us the hs_err_g1 files too. :-)
>>> Yes. Agreed.
>>> 
>>> Ship it!
>>> 
>>> Bengt
>>>>> 
>>>>>> 
>>>>>> 
>>>>>> 
>>>>>> I don't think the changes in the webrev are too controversial. The reason I wanted to give a heads up to both groups was to point out that this change will increase the size of the hs_err files when G1 is used. It's common to have 1,000 to 2,000 regions in the heap, if larger heaps are used (much fewer if smaller heaps are used), which means that the hs_err file will have this many extra lines. Does anyone have any concerns about this?
>>>>>> 
>>>>>> 
>>>>>> 
>>>>>> 



More information about the hotspot-gc-dev mailing list