Extremely long parnew/cms promotion failure scenario?
Srinivas Ramakrishna
ysr1729 at gmail.com
Tue Oct 30 09:49:14 PDT 2012
Hi Peter, all --
On Fri, Oct 19, 2012 at 1:40 AM, Srinivas Ramakrishna <ysr1729 at gmail.com>wrote:
>
>
> On Thu, Oct 18, 2012 at 5:27 PM, Peter B. Kessler <
> Peter.B.Kessler at oracle.com> wrote:
>
>> When there's no room in the old generation and a worker has filled its
>> PLAB to capacity, but it still has instances to try to promote, does it try
>> to allocate a new PLAB, and fail? That would lead to each of the workers
>> eventually failing to allocate a new PLAB for each promotion attempt.
>> IIRC, PLAB allocation grabs a real lock (since it happens so rarely :-).
>> In the promotion failure case, that lock could get incandescent. Maybe
>> it's gone unnoticed because for modest young generations it doesn't stay
>> hot enough for long enough for people to witness the supernova? Having a
>> young generation the size you do would exacerbate the problem. If you have
>> lots of workers, that would increase the amount of contention, too.
>>
>
> Yes, that's exactly my thinking too. For the case of CMS, the PLAB's are
> "local free block lists" and the allocation from the shared global pool is
> even worse and more heavyweight than an atomic pointer bump, with a lock
> protecting several layers of checks.
>
>
>>
>> PLAB allocation might be a place where you could put a test for having
>> failed promotion, so just return null and let the worker self-loop this
>> instance. That would keep the test off the fast-path (when things are
>> going well).
>>
>
> Yes, that's a good idea and might well be sufficient, and was also my
> first thought. However, I also wonder about whether just moving the
> promotion
> failure test a volatile read into the fast path of the copy routine, and
> immediately failing all subsequent copies after the first failure (and
> indeed via the
> global flag propagating that failure across all the workers immediately)
> won't just be quicker without having added that much in the fast path. It
> seems
> that in that case we may be able to even avoid the self-looping and the
> subsequent single-threaded fixup. The first thread that fails sets the
> volatile
> global, so any subsequent thread artificially fails all subsequent copies
> of uncopied objects. Any object reference found pointing to an object in
> Eden
> or From space that hasn't yet been copied will call the copy routine which
> will (artificially) fail and return the original address.
>
> I'll do some experiments and there may lurk devils in the details, but it
> seems to me that this will work and be much more efficient in the
> slow case, without making the fast path that much slower.
>
I got back to thinking about this issue with a view to implementing the
second approach above, but ran into what's an obvious problem which should
have occurred to me. We need a way of terminating the failure scan while at
the same time making sure to update all references that point at old copies
of objects that have already been copied.
We can do one or the other easily, but not both at the same time without
avoiding the linear walk of Eden and From spaces, which I had wanted to
avoid
on account of its single-threadedness.
(1) The simplest termination protocol is that once each thread sees that
promotion has failed, it does not attempt to copy the target object (if it
has not yet
been copied), but simply, returns the object from the copy routine.
It also does not push the object on its work stack. Thus the workers will
eventually
terminate their scan of the card-table and of the objects on their
work stacks after having updated any references to moved objects
encountered during
the remainder of the scan. This can however still leave a subset of
the references of objects in the young generation that were not copied
pointing to
the old copies of objects that were copied. Fixing those would require
a (linear) scan of Eden and From survivor spaces (which today is
single-threaded,
although it conceivably could be made multi-threaded by remembering
TLAB allocation boundaries and using those as the parallel chunks of work).
(2) If we didn't want to do the linear scan mentioned above, then at
termination, we should have scanned all reachable objects in the young gen.
In that
case, to terminate the scan we would have to remember when the
tracing had already scanned an object, so the trace would not need to
process that
object. Previously, this was done by looking at the location of the
object (not in Eden or From Space) or at its header (a forwarding pointer).
The use of
a self-loop forwarding pointer for failed copies requires another
pass to clear it which is the linear scan of Eden and From spaces that we
wanted to
avoid above. The self-looping also requires evacuation and the later
rematerialization of the header contents during the scan, both of which
carry
additional expense. Even if we had a spare header bit -- which we
well might in the 64-bit case, i think, although i haven't looked recently
-- we would
still need to clear that bit without doing a linear pass over Eden
and From spaces. We could conceivably do so as part of the later full gc
that is to follow
shortly, but that would add at least some bulk via a test to the full
gc closures.
So it almost seems as though the later scan of Eden and From spaces is
unavoidable, no matter what. Unless someone is able to think of a clever
way of avoiding it while still quickly terminating the tracing after
signalling global promotion failure.
Given that, my first though was to use the approach outlined in (1) where
we terminate quickly, then do a linear walk of Eden and From and fix up any
stale
references that were left pointing at old copies of successfully copied
objects. The advantage of that appears to be two-fold: firstly the algorithm
terminates the trace very quickly and does not pay any time or space
penalty of self-looping and header evacuation. The disadvantage however is
that the linear walk might need to do much more work in scanning all
objects and fixing their "stale" references, even when those objects may
themselves be dead.
To evaluate the rough relative tradeoff between the two schemes, one could
do the following back of the envelope calculation:
. Suppose that the object graph is random and that some small fraction s of
young gen objects y survive a scavenge on average.
. Then the work done in the linear walk of scheme (1) is O(y + e) where y
is the total number of objects, dead or alive, in Eden and From space,
and e is the total number of edges (references) in those objects.
. In the case of scheme (2), the work done in the linear walk is O(y) where
y is the total number of objects, dead or alive, in Eden and From space. The
space used and the work done for evacuating the non-prototypical headers
would be O(f.s.y) where s and y are as defined above, and f is the fraction
of survivors that have non-prototypical headers.
Also, the multiplicative constant factor in the big-O for scheme (1) seems
to be much higher than for scheme (2): while scheme (2) merely examines
the header of each object and fixes up the self-loop (or clears the header
bit if using a space header bit), scheme (1) needs to iterate over the
references
in each object, examine the header of the target object of that reference
and determine if the reference must be updated, and do so if needed.
Thus although (1) may terminate the trace very quickly, the linear scan of
Eden and From spaces for the "fix-up" phase is likely to be much more
expensive.
It could well be that (2) is the superior scheme under those circumstances,
which would bring us back to self-looping (or its moral equivalent a header
bit).
The header bit avoids the time and space expense of the spooling of
non-prototypical headers. If we optimized the acquisition of spooling space
for
headers, we might be almost on par with the header bit use and it would
leave the fast path of normal scavenge unaffected as is the case today.
(Use of the header bit would need an extra test.)
After thinking over this, it appears as though the simplest route is to do
what Peter mentioned earlier, which is to just have the allocation of the
promotion
buffers have fast-fail paths (after promotion failure) that do not take
locks, and just cut down on that expense.
Let me know if anyone has any comments/feedback/ideas: otherwise I'll go
through the CMS (and ParallelOld) allocation paths and place
"fast-fail gates" along the slow allocation path where we go to refill the
local buffers. (Later, we can also see about the possibility of using a
header bit
to signal a failed copy, rather than a self-loop, and see if it fetches any
gains by saving on header spooling while not affecting the fast path too
much,
and of multi-threading the "linear-scan" by using TLAB boundaries of the
previous epoch. But those would be later optimizations.)
-- ramki
>
>>
>> I'm still guessing.
>
>
> Your guesses are good, and very helpful, and I think we are on the right
> track with this one as regards the cause of the slowdown.
>
> I'll update.
>
> -- ramki
>
>
>>
>>
>> ... peter
>>
>> Srinivas Ramakrishna wrote:
>>
>>> System data show high context switching in vicinity of event and points
>>> at the futile allocation bottleneck as a possible theory with some legs....
>>>
>>> more later.
>>> -- ramki
>>>
>>> On Thu, Oct 18, 2012 at 3:47 PM, Srinivas Ramakrishna <ysr1729 at gmail.com<mailto:
>>> ysr1729 at gmail.com>> wrote:
>>>
>>> Thanks Peter... the possibility of paging or related issue of VM
>>> system did occur to me, especially because system time shows up as
>>> somewhat high here. The problem is that this server runs without
>>> swap :-) so the time is going elsewhere.
>>>
>>> The cache miss theory is interesting (but would not show up as
>>> system time), and your back of the envelope calculation gives about
>>> 0.8 us for fetching a cache line, although i am pretty sure the
>>> cache miss predictor would probably figure out the misses and stream
>>> in the
>>> cache lines since as you say we are going in address order). I'd
>>> expect it to be no worse than when we do an "initial mark pause on a
>>> full Eden", give or
>>> take a little, and this is some 30 x worse.
>>>
>>> One possibility I am looking at is the part where we self-loop. I
>>> suspect the ParNew/CMS combination running with multiple worker
>>> threads
>>> is hit hard here, if the failure happens very early say -- from what
>>> i saw of that code recently, we don't consult the flag that says we
>>> failed
>>> so we should just return and self-loop. Rather we retry allocation
>>> for each subsequent object, fail that and then do the self-loop. The
>>> repeated
>>> failed attempts might be adding up, especially since the access
>>> involves looking at the shared pool. I'll look at how that is done,
>>> and see if we can
>>> do a fast fail after the first failure happens, rather than try and
>>> do the rest of the scavenge, since we'll need to do a fixup anyway.
>>>
>>> thanks for the discussion and i'll update as and when i do some more
>>> investigations. Keep those ideas coming, and I'll submit a bug
>>> report once
>>> i have spent a few more cycles looking at the available data and
>>> ruminating.
>>>
>>> - ramki
>>>
>>>
>>> On Thu, Oct 18, 2012 at 1:20 PM, Peter B. Kessler
>>> <Peter.B.Kessler at oracle.com <mailto:Peter.B.Kessler@**oracle.com<Peter.B.Kessler at oracle.com>>>
>>> wrote:
>>>
>>> IIRC, promotion failure still has to finish the evacuation
>>> attempt (and some objects may get promoted while the ones that
>>> fail get self-looped). That part is the usual multi-threaded
>>> object graph walk, with failed PLAB allocations thrown in to
>>> slow you down. Then you get to start the pass that deals with
>>> the self-loops, which you say is single-threaded. Undoing the
>>> self-loops is in address order, but it walks by the object
>>> sizes, so probably it mostly misses in the cache. 40GB at the
>>> average object size (call them 40 bytes to make the math easy)
>>> is a lot of cache misses. How fast is your memory system?
>>> Probably faster than (10minutes / (40GB / 40bytes)) per cache
>>> miss.
>>>
>>> Is it possible you are paging? Maybe not when things are
>>> running smoothly, but maybe a 10 minute stall on one service
>>> causes things to back up (and grow the heap of) other services
>>> on the same machine? I'm guessing.
>>>
>>> ... peter
>>>
>>> Srinivas Ramakrishna wrote:
>>>
>>>
>>> Has anyone come across extremely long (upwards of 10
>>> minutes) promotion failure unwinding scenarios when using
>>> any of the collectors, but especially with ParNew/CMS?
>>> I recently came across one such occurrence with ParNew/CMS
>>> that, with a 40 GB young gen took upwards of 10 minutes to
>>> "unwind". I looked through the code and I can see
>>> that the unwinding steps can be a source of slowdown as we
>>> iterate single-threaded (DefNew) through the large Eden to
>>> fix up self-forwarded objects, but that still wouldn't
>>> seem to explain such a large pause, even with a 40 GB young
>>> gen. I am looking through the promotion failure paths to see
>>> what might be the cause of such a large pause,
>>> but if anyone has experienced this kind of scenario before
>>> or has any conjectures or insights, I'd appreciate it.
>>>
>>> thanks!
>>> -- ramki
>>>
>>>
>>> ------------------------------**
>>> __----------------------------**--__------------
>>>
>>> ______________________________**___________________
>>> hotspot-gc-use mailing list
>>> hotspot-gc-use at openjdk.java.__**net
>>> <mailto:hotspot-gc-use@**openjdk.java.net<hotspot-gc-use at openjdk.java.net>
>>> >
>>> http://mail.openjdk.java.net/_**
>>> _mailman/listinfo/hotspot-gc-_**_use<http://mail.openjdk.java.net/__mailman/listinfo/hotspot-gc-__use>
>>> <http://mail.openjdk.java.net/**mailman/listinfo/hotspot-gc-
>>> **use <http://mail.openjdk.java.net/mailman/listinfo/hotspot-gc-use>>
>>>
>>>
>>>
>>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/hotspot-gc-use/attachments/20121030/46ebba9b/attachment-0001.html
More information about the hotspot-gc-use
mailing list