RFR: 8357959: (bf) ByteBuffer.allocateDirect initialization can result in large TTSP spikes [v8]

[Maurizio Cimadamore]

On Tue, 3 Jun 2025 10:32:13 GMT, Rohitash Kumar <duke at openjdk.org> wrote:

>> ByteBuffer.allocateDirect uses UNSAFE.setMemory, causing high time-to-safepoint (100+ ms) for large (100 MB+) allocations.
>> 
>> This PR applies a simple fix by chunking the zeroing operation within ByteBuffers. A more robust solution would be to add chunking inside UNSAFE.setMemory itself. However Its not that straightforward as mentioned by Aleksey in [JDK-8357959](https://bugs.openjdk.org/browse/JDK-8357959)
>>>Looks like all current uses we care about are in Buffers. Taking a safepoint within cleaning would open some questions whether any VM code expect to see semi-initialized area we are busy cleaning up. For Buffers, this question does not arise. Therefore, we can do the fix in Buffers first, without changing the Unsafe itself.
>>  
>> I can pursue that if its preferred. I chose 1 MB as a chunk size some what arbitrarily I am open to suggestion, if there are better options.
>> 
>> For verification, I tested the fix against the reproducer - [gist](https://gist.github.com/rk-kmr/be4322b72a14ae04aeefc0260c01acf6) and confirmed that ttsp timing were lower.
>> 
>> **before**
>> 
>> 0.444s][info][safepoint,stats] ThreadDump                   [             13               1 ][        194156625      65291  194221916 ]               0
>> [0.662s][info][safepoint,stats] ThreadDump                   [             13               1 ][        200013875      87834  200101709 ]               0
>> [0.858s][info][safepoint,stats] ThreadDump                   [             13               1 ][        183762583      43417  183806000 ]               0
>> [1
>> 
>> **after**
>> 
>> 1.705s][info][safepoint,stats] ThreadDump                   [             11               1 ][            92792      24958     117750 ]               0
>> [1.724s][info][safepoint,stats] ThreadDump                   [             11               1 ][           497375      94041     591416 ]               0
>> [1.736s][info][safepoint,stats] ThreadDump                   [             11               1 ][           156750      47208     203958 ]               0
>> [1.747s][info][safepoint,stats] ThreadDump                   [             11               1 ][           121958      28334     150292 ]               0
>> 
>> 
>> I added a benchmark to ensure that chunking doesn't introduce significant overhead across different allocation sizes, and following results confirm that. 
>> 
>> **Before**
>> 
>> Benchmark                                              (bytes)  Mode  Cnt          Score         Error  Units
>> B...
>
> Rohitash Kumar has updated the pull request incrementally with two additional commits since the last revision:
> 
>  - rename param from size to count
>  - Apply patch for test extension

Note: in FFM we actually use a plain `long` loop instead of calling `Unsafe::setMemory`:
https://github.com/openjdk/jdk/blob/master/src/java.base/share/classes/jdk/internal/foreign/SegmentFactories.java#L239

(to facilitate this, we ensure that all allocations are, internally, at least 8-byte aligned).

A simple loop like this, with only one input region of memory typically auto-vectorizes quite well. In the long run, I'd prefer to see bulk operations in Unsafe being rewritten to use simpler operations and loops. I think we're getting quite close, but we can't get all the way there for stuff like `copyMemory` because, currently, there are some problems in determining that two off-heap memory regions are provably disjoint (but this is not an issue in the case of `setMemory`).

-------------

PR Comment: https://git.openjdk.org/jdk/pull/25487#issuecomment-2940835094