Free ratio based heap shrinking in the parallel collector

[Hiroshi Yamauchi]

On Tue, Mar 16, 2010 at 3:38 PM, Jon Masamitsu <Jon.Masamitsu at sun.com>wrote:

> Hiroshi,
>
> The goal of UseAdaptiveSizePolicy with regard to performance is
> to size the generations so as to achieve the throughput goal.
> GC cost is reduced by reducing the frequency of collections
> and that is achieved by more free space in the generations
> (more space for more allocations before the next GC).
>

OK.


>
> I can see how a user might want to limit the footprint.  Without
> your change I would tell the user to reduce the throughput goal.
> Not a particularly friendly interface to limit footprint I'll
> admit.
>

The motive is to make it easier for users to control the footprint. The free
ratio is very easy to understand for users.


>
> There is a method
>
> PSAdaptiveSizePolicy::adjust_for_throughput()
>
> which is called to calculate the amount that the heap
> should grow in order to achieve a throughput goal.
> If that method said not to increase the heap based
> on the value of MaxHeapFreeRatio is that close to
> what you want?  It's not the same but it would be
> a way of limiting the growth of the heap (stops
> growing the heap when the free space after the collection
> exceeds a threshold).  I think it would fit in more
> easily with the rest of UseAdativeSizePolicy.  I think
> it would be easier to explain too.
>

I think even if adjust_for_throughput tries to limit the growth that way in
each invocation of it, it's likely that the heap will eventually fully
expanded with enough amount of load (and objects). So, it could delay the
expansion under light load. But I think the key here is to be able to shrink
the heap after it is expanded. It looks to me that adjust_for_throughput
never shrinks the size. If I'm not mistaken, it's closer, but not quite
there, I think.

Hiroshi


> Jon
>
> On 03/15/10 16:50, Hiroshi Yamauchi wrote:
>
>
>>
>> On Mon, Mar 15, 2010 at 12:03 PM, Jon Masamitsu <Jon.Masamitsu at sun.com<mailto:
>> Jon.Masamitsu at sun.com>> wrote:
>>
>>    Hiroshi,
>>
>>    I'm looking at the code in psParallelCompact.cpp
>>
>>    2076     bool free_ratio_in_effect = false;
>>    2077     if ((UseFreeRatioForParallelGC ||
>>    2078          (UseFreeRatioOnlyInSystemGCForParallelGC &&
>>    2079           gc_cause == GCCause::_java_lang_system_gc))) {
>>    2080       ParallelScavengeHeap* heap =
>>    (ParallelScavengeHeap*)Universe::heap();
>>    2081       free_ratio_in_effect =
>>    heap->try_to_shrink_by_free_ratio(true);
>>    2082     }
>>    2083
>>    2084     if (!free_ratio_in_effect && UseAdaptiveSizePolicy) {
>>
>>    If only UseFreeRatioForParallelGC is set to true, is it correct
>>    that the heap
>>
>>     a) will grow according to the original UseAdaptiveSizePolicy
>>    policy until MaxHeapFreeRatio is exceeded (i.e., until
>>    try_to_shrink_by_free_ratio() returns true)
>>     b) will then shrink until MaxHeapFreeRatio is no longer exceeded
>>    (likely just that once)
>>
>>    then start with a) again?
>>
>>
>> I think that mostly matches my intention. The cyclical pattern
>> (a->b->a->b->...) is assumed not to happen more often than necessary though.
>> An expectation is that the value of MaxHeapFreeRatio is high enough (70 by
>> default) that the b) only happens when the app is under reasonably light(er)
>> load (or just idle.) UseFreeRatioOnlyInSystemGCForParallelGC (instead of
>> UseFreeRatioForParallelGC ) is for apps that call System.gc() to give a hint
>> about the need to have its heap shrunk and want to avoid heap shrinking
>> otherwise.
>>
>> An intuition is the following. To reduce the memory footprint, the free
>> ratio logic holds back the heap expansion if the original adaptive size
>> policy expands too aggressively, or overrides and shrink the heap if the
>> original adaptive size policy leaves too much free space, in terms of the
>> value of MaxHeapFreeRatio. I think it's not far from what flag
>> MaxHeapFreeRatio originally meant. Otherwise, the heap size depends on the
>> original adaptive size policy (for performance.) It's a tradeoff between
>> performance and footprint.
>>
>>
>>    What type of application fits with such a combination of policies?
>>    Do you have a plots of how the heap sizing behaves (i.e., size of the
>>    generations vs. collections/time)?
>>
>>
>> The reason to combine the two logics (size policies) isn't for particular
>> applications. I don't get too surprised if you say it should be a separate
>> size policy and leave the original adaptive size policy alone. My thinking
>> is that users may want the same level of performance from the parallel
>> collector (the adaptive size policy is enabled by default) unless a large
>> part of the heap (> 70% by default) is free. This is my attempt to get the
>> best of both. The flags are off by default since of course, not all
>> applications need to have their heap shrunk this way.
>>
>> No, I don't have plots.
>>
>> Hiroshi
>>
>>
>>
>>    Jon
>>
>>    Hiroshi Yamauchi wrote On 03/10/10 15:17,:
>>
>>     > Hi,
>>     >
>>     > I'd like to have this patch
>>     >
>>     >   http://cr.openjdk.java.net/~rasbold/69XXXXX/webrev.00/<http://cr.openjdk.java.net/%7Erasbold/69XXXXX/webrev.00/>
>>    <http://cr.openjdk.java.net/%7Erasbold/69XXXXX/webrev.00/>
>>     > <http://cr.openjdk.java.net/%7Erasbold/69XXXXX/webrev.00/>
>>     >
>>     > contributed, if appropriate.
>>     >
>>     > It implements a simple heap shrinking logic based on the free ratio
>>     > (MaxHeapFreeRatio) in the parallel collector. The way it works is
>> the
>>     > following: If the free ratio, (1.0 - <used> / <capacity>) * 100 >
>>     > MaxHeapFreeRatio, this logic kicks in. Otherwise, the (default)
>>     > adaptive size policy runs the show (as before). This feature is
>>    turned
>>     > off by default. There are two new flags to turn it on:
>>     >
>>     >   UseFreeRatioForParallelGC - this enables it.
>>     >   UseFreeRatioOnlyInSystemGCForParallelGC - this enables it only in
>>     > explicit System.gc() calls. This is more useful in apps that call
>>     > System.gc() when it's idle and don't want shrinking to happen at all
>>     > other times.
>>     >
>>     > In our tests, it appears to work well in a simple test and to help a
>>     > real app reduce its footprint (RSS) when it's idle.
>>     >
>>     > Thanks to Chuck Rasbold who uploaded the webrev on my behalf.
>>     >
>>     > Thanks,
>>     > Hiroshi
>>     >
>>
>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/hotspot-gc-dev/attachments/20100318/e7697fbe/attachment.htm>