RFR(L): 8198423: Improve metaspace chunk allocation (was: Proposal for improvements to the metaspace chunk allocator)

Erik Helin erik.helin at oracle.com
Thu Mar 8 16:14:36 UTC 2018

Alright, I know I am bit late to the party, but I have now refreshed my 
knowledge of Metaspace enough to begin reviewing this. I will probably 
ask a few questions to get a better understanding of your changes :)

First of all, great work! The code is easy to read and the concepts are 
clear, so thank you for that.

An initial commend:

1592     // Chunks are born as in-use (see MetaChunk ctor). So, before 
1593     // the padding chunk to its chunk manager, mark it as in use 
1594     // will assert that).
1595     do_update_in_use_info_for_chunk(padding_chunk, true);

This comment is slightly hard to read. I _think_ what you are meaning is 
something like:

    // So, before returning the padding chunk to its chunk manger,
    // mark it as in use in the the occupancy map.

Is that correct? Besides updating the occupancy map, 
do_update_in_use_info_for_chunk will also call 
MetaChunk::set_is_tagged_free, but that is not strictly needed here, right?


On 02/28/2018 05:17 PM, Thomas Stüfe wrote:
> Hi Eric, no problem!
> Thanks, Thomas
> On Wed, Feb 28, 2018 at 4:28 PM, Erik Helin <erik.helin at oracle.com 
> <mailto:erik.helin at oracle.com>> wrote:
>     Hi Thomas,
>     I will take a look at this, I just have been a bit busy lately
>     (sorry for not responding earlier).
>     Thanks,
>     Erik
>     On 02/26/2018 03:20 PM, Thomas Stüfe wrote:
>         Hi all,
>         I know this patch is a bit larger, but may I please have reviews
>         and/or
>         other input?
>         Issue: https://bugs.openjdk.java.net/browse/JDK-8198423
>         <https://bugs.openjdk.java.net/browse/JDK-8198423>
>         Latest version:
>         http://cr.openjdk.java.net/~stuefe/webrevs/metaspace-coalescation/2018-02-26/webrev/
>         <http://cr.openjdk.java.net/~stuefe/webrevs/metaspace-coalescation/2018-02-26/webrev/>
>         For those who followed the mail thread, this is the incremental
>         diff to the
>         last changes (included feedback Goetz gave me on- and off-list):
>         http://cr.openjdk.java.net/~stuefe/webrevs/metaspace-coalescation/2018-02-26/webrev-incr/webrev/
>         <http://cr.openjdk.java.net/~stuefe/webrevs/metaspace-coalescation/2018-02-26/webrev-incr/webrev/>
>         Thank you!
>         Kind Regards, Thomas Stuefe
>         On Thu, Feb 8, 2018 at 12:58 PM, Thomas Stüfe
>         <thomas.stuefe at gmail.com <mailto:thomas.stuefe at gmail.com>>
>         wrote:
>             Hi,
>             We would like to contribute a patch developed at SAP which
>             has been live
>             in our VM for some time. It improves the metaspace chunk
>             allocation:
>             reduces fragmentation and raises the chance of reusing free
>             metaspace
>             chunks.
>             The patch:
>             http://cr.openjdk.java.net/~stuefe/webrevs/metaspace-coalesc
>             <http://cr.openjdk.java.net/~stuefe/webrevs/metaspace-coalesc>
>             ation/2018-02-05--2/webrev/
>             In very short, this patch helps with a number of
>             pathological cases where
>             metaspace chunks are free but cannot be reused because they
>             are of the
>             wrong size. For example, the metaspace freelist could be
>             full of small
>             chunks, which would not be reusable if we need larger
>             chunks. So, we could
>             get metaspace OOMs even in situations where the metaspace
>             was far from
>             exhausted. Our patch adds the ability to split and merge
>             metaspace chunks
>             dynamically and thus remove the "size-lock-in" problem.
>             Note that there have been other attempts to get a grip on
>             this problem,
>             see e.g. "SpaceManager::get_small_chunks_and_allocate()".
>             But arguably
>             our patch attempts a more complete solution.
>             In 2016 I discussed the idea for this patch with some folks
>             off-list,
>             among them Jon Matsimutso. He then did advice me to create a
>             JEP. So I did:
>             [1]. However, meanwhile changes to the JEP process were
>             discussed [2], and
>             I am not sure anymore this patch needs even needs a JEP. It
>             may be
>             moderately complex and hence carries the risk inherent in
>             any patch, but
>             its effects would not be externally visible (if you discount
>             seeing fewer
>             metaspace OOMs). So, I'd prefer to handle this as a simple RFE.
>             --
>             How this patch works:
>             1) When a class loader dies, its metaspace chunks are freed
>             and returned
>             to the freelist for reuse by the next class loader. With the
>             patch, upon
>             returning a chunk to the freelist, an attempt is made to
>             merge it with its
>             neighboring chunks - should they happen to be free too - to
>             form a larger
>             chunk. Which then is placed in the free list.
>             As a result, the freelist should be populated by larger
>             chunks at the
>             expense of smaller chunks. In other words, all free chunks
>             should always be
>             as "coalesced as possible".
>             2) When a class loader needs a new chunk and a chunk of the
>             requested size
>             cannot be found in the free list, before carving out a new
>             chunk from the
>             virtual space, we first check if there is a larger chunk in
>             the free list.
>             If there is, that larger chunk is chopped up into n smaller
>             chunks. One of
>             them is returned to the callers, the others are re-added to
>             the freelist.
>             (1) and (2) together have the effect of removing the
>             size-lock-in for
>             chunks. If fragmentation allows it, small chunks are
>             dynamically combined
>             to form larger chunks, and larger chunks are split on demand.
>             --
>             What this patch does not:
>             This is not a rewrite of the chunk allocator - most of the
>             mechanisms stay
>             intact. Specifically, chunk sizes remain unchanged, and so
>             do chunk
>             allocation processes (when do which class loaders get handed
>             which chunk
>             size). Almost everthing this patch does affects only
>             internal workings of
>             the ChunkManager.
>             Also note that I refrained from doing any cleanups, since I
>             wanted
>             reviewers to be able to gauge this patch without filtering
>             noise.
>             Unfortunately this patch adds some complexity. But there are
>             many future
>             opportunities for code cleanup and simplification, some of
>             which we already
>             discussed in existing RFEs ([3], [4]). All of them are out
>             of the scope for
>             this particular patch.
>             --
>             Details:
>             Before the patch, the following rules held:
>             - All chunk sizes are multiples of the smallest chunk size
>             ("specialized
>             chunks")
>             - All chunk sizes of larger chunks are also clean multiples
>             of the next
>             smaller chunk size (e.g. for class space, the ratio of
>             specialized/small/medium chunks is 1:2:32)
>             - All chunk start addresses are aligned to the smallest
>             chunk size (more
>             or less accidentally, see metaspace_reserve_alignment).
>             The patch makes the last rule explicit and more strict:
>             - All (non-humongous) chunk start addresses are now aligned
>             to their own
>             chunk size. So, e.g. medium chunks are allocated at
>             addresses which are a
>             multiple of medium chunk size. This rule is not extended to
>             humongous
>             chunks, whose start addresses continue to be aligned to the
>             smallest chunk
>             size.
>             The reason for this new alignment rule is that it makes it
>             cheap both to
>             find chunk predecessors of a chunk and to check which chunks
>             are free.
>             When a class loader dies and its chunk is returned to the
>             freelist, all we
>             have is its address. In order to merge it with its neighbors
>             to form a
>             larger chunk, we need to find those neighbors, including
>             those preceding
>             the returned chunk. Prior to this patch that was not easy -
>             one would have
>             to iterate chunks starting at the beginning of the
>             VirtualSpaceNode. But
>             due to the new alignment rule, we now know where the
>             prospective larger
>             chunk must start - at the next lower
>             larger-chunk-size-aligned boundary. We
>             also know that currently a smaller chunk must start there (*).
>             In order to check the free-ness of chunks quickly, each
>             VirtualSpaceNode
>             now keeps a bitmap which describes its occupancy. One bit in
>             this bitmap
>             corresponds to a range the size of the smallest chunk size
>             and starting at
>             an address aligned to the smallest chunk size. Because of
>             the alignment
>             rules above, such a range belongs to one single chunk. The
>             bit is 1 if the
>             associated chunk is in use by a class loader, 0 if it is free.
>             When we have calculated the address range a prospective
>             larger chunk would
>             span, we now need to check if all chunks in that range are
>             free. Only then
>             we can merge them. We do that by querying the bitmap. Note
>             that the most
>             common use case here is forming medium chunks from smaller
>             chunks. With the
>             new alignment rules, the bitmap portion covering a medium
>             chunk now always
>             happens to be 16- or 32bit in size and is 16- or 32bit
>             aligned, so reading
>             the bitmap in many cases becomes a simple 16- or 32bit load.
>             If the range is free, only then we need to iterate the
>             chunks in that
>             range: pull them from the freelist, combine them to one new
>             larger chunk,
>             re-add that one to the freelist.
>             (*) Humongous chunks make this a bit more complicated. Since
>             the new
>             alignment rule does not extend to them, a humongous chunk
>             could still
>             straddle the lower or upper boundary of the prospective
>             larger chunk. So I
>             gave the occupancy map a second layer, which is used to mark
>             the start of
>             chunks.
>             An alternative approach could have been to make humongous
>             chunks size and
>             start address always a multiple of the largest non-humongous
>             chunk size
>             (medium chunks). That would have caused a bit of waste per
>             humongous chunk
>             (<64K) in exchange for simpler coding and a simpler
>             occupancy map.
>             --
>             The patch shows its best results in scenarios where a lot of
>             smallish
>             class loaders are alive simultaneously. When dying, they
>             leave continuous
>             expanses of metaspace covered in small chunks, which can be
>             merged nicely.
>             However, if class loader life times vary more, we have more
>             interleaving of
>             dead and alive small chunks, and hence chunk merging does
>             not work as well
>             as it could.
>             For an example of a pathological case like this see example
>             program: [5]
>             Executed like this: "java -XX:CompressedClassSpaceSize=10M
>             -cp test3
>             test3.Example2" the test will load 3000 small classes in
>             separate class
>             loaders, then throw them away and start loading large
>             classes. The small
>             classes will have flooded the metaspace with small chunks,
>             which are
>             unusable for the large classes. When executing with the
>             rather limited
>             CompressedClassSpaceSize=10M, we will run into an OOM after
>             loading about
>             800 large classes, having used only 40% of the class space,
>             the rest is
>             wasted to unused small chunks. However, with our patch the
>             example program
>             will manage to allocate ~2900 large classes before running
>             into an OOM, and
>             class space will show almost no waste.
>             Do demonstrate this, add -Xlog:gc+metaspace+freelist. After
>             running into
>             an OOM, statistics and an ASCII representation of the class
>             space will be
>             shown. The unpatched version will show large expanses of
>             unused small
>             chunks, the patched variant will show almost no waste.
>             Note that the patch could be made more effective with a
>             different size
>             ratio between small and medium chunks: in class space, that
>             ratio is 1:16,
>             so 16 small chunks must happen to be free to form one larger
>             chunk. With a
>             smaller ratio the chance for coalescation would be larger.
>             So there may be
>             room for future improvement here: Since we now can merge and
>             split chunks
>             on demand, we could introduce more chunk sizes. Potentially
>             arriving at a
>             buddy-ish allocator style where we drop hard-wired chunk
>             sizes for a
>             dynamic model where the ratio between chunk sizes is always
>             1:2 and we
>             could in theory have no limit to the chunk size? But this is
>             just a thought
>             and well out of the scope of this patch.
>             --
>             What does this patch cost (memory):
>                - the occupancy bitmap adds 1 byte per 4K metaspace.
>                - MetaChunk headers get larger, since we add an enum and
>             two bools to it.
>             Depending on what the c++ compiler does with that, chunk
>             headers grow by
>             one or two MetaWords, reducing the payload size by that amount.
>             - The new alignment rules mean we may need to create padding
>             chunks to
>             precede larger chunks. But since these padding chunks are
>             added to the
>             freelist, they should be used up before the need for new
>             padding chunks
>             arises. So, the maximally possible number of unused padding
>             chunks should
>             be limited by design to about 64K.
>             The expectation is that the memory savings by this patch far
>             outweighs its
>             added memory costs.
>             .. (performance):
>             We did not see measurable drops in standard benchmarks
>             raising over the
>             normal noise. I also measured times for a program which
>             stresses metaspace
>             chunk coalescation, with the same result.
>             I am open to suggestions what else I should measure, and/or
>             independent
>             measurements.
>             --
>             Other details:
>             I removed SpaceManager::get_small_chunk_and_allocate() to reduce
>             complexity somewhat, because it was made mostly obsolete by
>             this patch:
>             since small chunks are combined to larger chunks upon return
>             to the
>             freelist, in theory we should not have that many free small
>             chunks anymore
>             anyway. However, there may be still cases where we could
>             benefit from this
>             workaround, so I am asking your opinion on this one.
>             About tests: There were two native tests -
>             ChunkManagerReturnTest and
>             TestVirtualSpaceNode (the former was added by me last year)
>             - which did not
>             make much sense anymore, since they relied heavily on
>             internal behavior
>             which was made unpredictable with this patch.
>             To make up for these lost tests,  I added a new gtest which
>             attempts to
>             stress the many combinations of allocation pattern but does
>             so from a layer
>             above the old tests. It now uses Metaspace::allocate() and
>             friends. By
>             using that point as entry for tests, I am less dependent on
>             implementation
>             internals and still cover a lot of scenarios.
>             --
>             Review pointers:
>             Good points to start are
>             - ChunkManager::return_single_chunk() - specifically,
>             ChunkManager::attempt_to_coalesce_around_chunk() - here we
>             merge chunks
>             upon return to the free list
>             - ChunkManager::free_chunks_get(): Here we now split large
>             chunks into
>             smaller chunks on demand
>             - VirtualSpaceNode::take_from_committed() : chunks are allocated
>             according to align rules now, padding chunks are handles
>             - The OccupancyMap class is the helper class implementing
>             the new
>             occupancy bitmap
>             The rest is mostly chaff: helper functions, added tests and
>             verifications.
>             --
>             Thanks and Best Regards, Thomas
>             [1] https://bugs.openjdk.java.net/browse/JDK-8166690
>             <https://bugs.openjdk.java.net/browse/JDK-8166690>
>             [2]
>             http://mail.openjdk.java.net/pipermail/jdk-dev/2017-November
>             <http://mail.openjdk.java.net/pipermail/jdk-dev/2017-November>
>             /000128.html
>             [3] https://bugs.openjdk.java.net/browse/JDK-8185034
>             <https://bugs.openjdk.java.net/browse/JDK-8185034>
>             [4] https://bugs.openjdk.java.net/browse/JDK-8176808
>             <https://bugs.openjdk.java.net/browse/JDK-8176808>
>             [5]
>             https://bugs.openjdk.java.net/secure/attachment/63532/test3.zip
>             <https://bugs.openjdk.java.net/secure/attachment/63532/test3.zip>

More information about the hotspot-dev mailing list