RFR(xs): 8152118: MinTLABSize should be less than TLAB max

sangheon sangheon.kim at oracle.com
Fri Mar 18 17:28:21 UTC 2016


Hi Bengt,

Thanks for reviewing this.

On 03/18/2016 12:15 AM, Bengt Rutisson wrote:
>
> Hi Sangheon,
>
> On 2016-03-18 06:43, sangheon wrote:
>> Hi all,
>>
>> Could I have a couple of reviews for this tiny change?
>>
>> MinTLABSize doesn't have upper limit now and this change is proposing 
>> to be less than or equal to ThreadLocallAllocBuffer::max_size().
>>
>> CR: https://bugs.openjdk.java.net/browse/JDK-8152118
>> Webrev: http://cr.openjdk.java.net/~sangheki/8152118/webrev.00/
>
> I think this looks good.
Thanks.

>
> However, I have a related question. When I set up the min tlab size 
> high I get some issues reported from other constraint functions:
>
> $ java  -XX:MinTLABSize=16m -XX:G1HeapRegionSize=32m -version
> YoungPLABSize (4096) must be greater than or equal to ergonomic PLAB 
> minimum size (2097152)
> OldPLABSize (1024) must be greater than or equal to ergonomic PLAB 
> minimum size (2097152)
> Error: Could not create the Java Virtual Machine.
> Error: A fatal exception has occurred. Program will exit.
>
> I don't think this is a problem with the patch that you are proposing 
> now, but I thought I should mention it. Do you think this is something 
> we would like to fix in the argument parsing (like adjusting the PLAB 
> max to be at least PLAB min) or should we just let the users set up 
> the PLAB sizes if they use large TLAB sizes?
>
> This is a pre-exiting problem. Even without your patch I get the same 
> behavior. So, you can consider your proposed patched reviewed.
Right, we have this problem when we introduced range/constraint for 
command-line flags (JDK-8059557: JEP 245: Validate JVM Command-Line Flag 
Arguments).
And adjusting the flag during argument processing was out-of-scope. 
(Non-goal: second item)
Your opinion seems like adjusting related flags, am I correct? i.e. We 
set 'MinTLABSize' but the error happened related flags(YoungPLABSize and 
OldPLABSize). And then, I think it would be better to let the users 
set-up them because it will be hard to adjusting related flags. What 
extends we would decide candidate flags to adjust.

But I agree with you that it would be annoying for the users.

Thanks,
Sangheon


>
> Thanks,
> Bengt
>
>
>> Testing: JPRT, TestOptionsWithRanges.java for all platforms.
>>
>> Thanks,
>> Sangheon
>




More information about the hotspot-gc-dev mailing list