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

Mon Feb 26 14:20:58 UTC 2018

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
Latest version:
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/

Thank you!

Kind Regards, Thomas Stuefe

On Thu, Feb 8, 2018 at 12:58 PM, Thomas Stüfe <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
> 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
> [2] http://mail.openjdk.java.net/pipermail/jdk-dev/2017-November
> /000128.html
> [3] https://bugs.openjdk.java.net/browse/JDK-8185034
> [4] https://bugs.openjdk.java.net/browse/JDK-8176808
> [5] https://bugs.openjdk.java.net/secure/attachment/63532/test3.zip
>
>
>