Dmitry Konstantinov

Interested in Java performance tuning, distributed systems and high availability

[LaurieWired @lauriewired]

Modern DRAM is based on a brilliant design from IBM.

But, we're still paying for a latency penalty that's existed since the 60s!

In this video, I'm introducing my research project (Tailslayer) that immensely reduces p99.99 latency on traditional RAM!

By implementing a hedged read strategy taking advantage of (undocumented!) channel scrambling offsets, I've gotten as much as 15x reductions in tail latency.

The technique works across Intel, AMD, Graviton, DDR4, DDR5, x86, ARM, you name it.

Check out the C++ lib I wrote, watch the video, and try it yourself!

[Nitsan "Yak" Wakart @nitsanw]

It still has to be reviewed, but finally the cursor based C* compaction patch is in the open: https://t.co/SXN33mct4p
"3-5x faster [compaction] in most scenarios and allocates ~20mb vs. multiple GB[per compaction]"

:-)

[Dmitry Konstantinov @java_dnk]

@usehex я в детстве такую машину видел вживую в типографии, печатаешь текст и строчки отливаются в металле, это вам не твиты штамповать 😎

[Ionut Balosin @ionutbalosin]

🚀 Call for Contributors

If you're interested in contributing to the #JVM #Performance #Benchmarks project - an initiative that gained significant traction in the #Java community through our recent #JDK17 and #JDK11 analyses - check out the repo:

🔗 https://t.co/UozhLXVSo3

[Vivek Galatage @vivekgalatage]

A fantastic blog series on memory subsystem optimizations!!!

https://t.co/i9ee4oLrhP

[Dmitry Konstantinov @java_dnk]

@Nekrolm Байку про Математический мост без гвоздей гид тоже рассказал? :-)

[JEP Updates @Jepupdates]

JEP 509 ("JFR CPU-Time Profiling (Experimental)") added. See https://t.co/COVRI877sO

[Dmitry Vyukov @dvyukov]

Profiling: you are doing it all WRONG! The SHOCKING truth about CPU vs. real time

Now that I get your attention :)

My perf latency profiling changes merged:
https://t.co/D5713CXKEw

First profiler ever that samples realtime (not CPU time) and suitable for latency optimization

[Jean-Philippe Bempel @jpbempel.bsky.social @jpbempel]

Async-Profiler v4.0
Heatmaps and Native memory profiling and VMStructs-based stack walker
https://t.co/DVUba3aTZR

[Norman Maurer @normanmaurer]

I am excited to finally announce the release of #netty https://t.co/5pm9XCO5H4 which is in the making for quite some time! This release comes with a lot of new exciting features, including the highly anticipated #io_uring support. See https://t.co/dRuUUG5p9y

[Jaromir Hamala @jerrinot]

Interesting issue in OpenJDK's UUID.compareTo(): It uses signed comparisons, which leads to unexpected sorting when the highest bit in either word is set. It can't be fixed since altering a sort order between versions is a big no-no.

https://t.co/QlbTt0CD3x

[Dmitry Konstantinov @java_dnk]

@usehex @justphotonataly Стоит тогда и Rashomon посмотреть, можно сказать детектив, несколько версий о том, что случилось (чем-то напоминает китайский Герой, в плане структуры сюжета)

[Dmitry Konstantinov @java_dnk]

@Nekrolm Чай с молоком надо, для полноты 🙂

[Roy van Rijn @royvanrijn]

🚨 There has been a breakthrough in the world of hash tables 🚨

A paper that disproves the longstanding conjecture that uniform hashing is optimal.

"Optimal Bounds for Open Addressing Without Reordering" by Martín Farach-Colton, Andrew Krapivin, William Kuszmaul

https://t.co/srRzqWlXFe

They demonstrate it is possible to construct an open-addressed hash table with improved expected search complexities, both amortized and worst-case, without the need to reorder elements over time.

I've implemented a crude version in Java and the results indeed seem promising, are there any initiatives to have such a hash table option in Java?

I'd love to test this further, with real world examples, the code of course is far from perfect and optimized, while Java's implementation has been maximally tweaked and refined.

[Arpit Bhayani @arpit_bhayani]

a university undergrad made hash tables faster and broke 40-year-old belief 🤯

for 40 years, it was believed that we cannot have better than O(n) operation for hash table worst case inserts and searches if it uses open addressing - for the hash index, look linearly until you find your element or an empty slot to insert.

the paper published last month showed that we can construct a hash table that achieves far better search complexities (both average and worst-case) of (log n)^2 - better than n.

I have not yet read the paper, but if you are interested in digging deeper, I have linked the paper and the write-up in a subsequent tweet.

in case you are interested in knowing the hash table internals, I have an in-depth playlist talking about how they are implemented internally and why they are the way they are. again link in a subsequent tweet.

[Roman Kennke 🗽 @rkennke]

Lilliput 2 (aka 4-byte-object-headers) is now in the Lilliput project repo (based on JDK25-dev). For testing, grab nightlies from here (thanks, @shipilev):
https://t.co/bAguegfmpb

[Andy Pavlo (@andypavlo.bsky.social) @andy_pavlo]

New @CMUDB Course: Database Query Optimization!
A survey of how to build a SQL optimizer from scratch covering foundational and state-of-the-art implementations. All lectures available on Youtube: https://t.co/I8dMflK9dP

[Jens Axboe @axboe]

https://t.co/905KYtIivG

io_uring for Java.

"Doing Random reads JUring achieves 33% better performance than Java NIO FileChannel operations for local files and 78% better performance for remote files."

[Dmitry Konstantinov @java_dnk]

@FranckPachot An article why it can be not a good idea in some cases: https://t.co/GzRh75wgkj

[Aleksey Shipilëv @shipilev]

Parts of Leyden and other CDS improvements are now in JDK mainline, in time for JDK 24 fork. The improvements are visible even on simple HelloWorld-style examples. There are even more things to squeeze!