Making stuff fast at @redpandadata.
Canada
Joined December 2018
- Tweets12.6K
- Following176
- Followers4.2K
- Likes5.7K
Pinned Tweet
Do you like footnotes? Do you like long, rambling lists?
Then, this might be the thing for you:
travisdowns.github.io/blog/2…
There, I smash my personal footnote count record (20). As for lists? It has lists *within* lists. The only thing left is to write it in LISP.
Travis Downs retweeted
The latest release of the simdutf C++ library (6.0.0) brings in more convenient for C++20 users. While you used to have to provide both a pointer and a size parameter... often you can now just pass your container...
std::vector data{1, 2, 3, 4, 5};
// C++11 API
auto cpp11 = simdutf::autodetect_encoding(data.data(), data.size());
// C++20 API
auto cpp20 = simdutf::autodetect_encoding(data);
Link in the comments.
Travis Downs retweeted
to our friends/companies downloading #redpanda 1.5M times a day... should turn on caching 😂
So we have LLMs that border on indsiguishable from real people (well, this depends in part on the company you keep) but autocorrect on my phone is still awful, making mistakes a toddler could correct.
Is it a latency problem?
Travis Downs retweeted
We're excited to hear @redpandadata's Travis Downs describe the practical experience of building high performance systems with C++20 in an asynchronous runtime. He'll also discuss tradeoffs in adopting a thread-per-core architecture. bit.ly/43RHzlV
#P99CONF #ScyllaDB
Despite the expansive title this is mostly a quick look at coroutine performance for those who are interested in that kind of thing.
📣 WATCH REDPANDA AT #P99CONF
Join @trav_downs this Oct 19th to learn all about our adventures in thread-per-core async with Redpanda and #Seastar.
It's a virtual event, so all you have to do is register👇
p99conf.io/
Travis Downs retweeted
For a long time I've been thinking that using a closed loop (sync) for measuring latency is wrong
Travis Downs retweeted
Measuring the memory usage of your C++ program lemire.me/blog/2022/11/10/me…
Travis Downs retweeted
some pretty cool tests we are doing on the #redpanda cloud of aggregate throughputs of 40GB/s. nbd .... more deets to come on Nov 15th. hopin.com/events/redpanda-op…
thanks to @trav_downs for the goodie you see below.... maybe we should do a 200GB/s test next? 🤣
#kafka
There is always something special about seeing a 30GB/s #redpanda cluster (10GB/s producer writes + 20GB/s for replicas). Very cool 😎
Travis Downs retweeted
10-20GB/s yes. We don't have anyone at 40GB/s yet. 400GbE not yet, but DM me if you want to test. This is the reason we have a dedicated Performance Engineering team
xitter.cf/trav_downs/status/1575…
Do you count cycles in your sleep?
Do you sweat over those last few μsecs?
Do you both love and hate Nagle's algorithm?
Join me in making future of event streaming wicked fast!
@emaxerrno has built a team of great people here at Redpanda and we're looking for a perf expert.
Travis Downs retweeted
New ARM instructions were just released! (With SVE 2.1, among many other things)
(I always use the exploration tools, but you can view them online too: developer.arm.com/documentat…)
developer.arm.com/downloads/…
Do you count cycles in your sleep?
Do you sweat over those last few μsecs?
Do you both love and hate Nagle's algorithm?
Join me in making future of event streaming wicked fast!
@emaxerrno has built a team of great people here at Redpanda and we're looking for a perf expert.
You can find additional details and apply here:
boards.greenhouse.io/redpand…
Alex's (@emaxerrno) and my DMs are always open for any questions, or (in my case) just random performance chit-chat is well received too.
Travis Downs retweeted
While everyone is enjoying all the #AMD #Zen4 performance reviews, here is my teardown of Zen4's #AVX512 implementation and architecture.
tl;dr - Intel has some serious competition now.
mersenneforum.org/showthread…
Travis Downs retweeted
y-cruncher v0.7.10 coming out soon with Zen4 AVX512 optimizations.
This was fun since there are no optimization resources yet for this chip. Had to do all the RE myself.😅
Travis Downs retweeted
New blog post: Faster zlib/DEFLATE decompression on the Apple M1 (and x86)
dougallj.wordpress.com/2022/…
ALT Graph of decompression performance, showing zlib-dougallj outperforming 3 alternatives by about 50% on each of the files in the Silesia Corpus. zlib-cloudflare is consistently second best, closely followed by zlib-ng, with zlib-apple further behind, outperformed by about 2x. The files have been compressed with zlib level 6, and throughput is measured in compressed mb/s on M1 using Apple clang 13.1.6. zlib-dougallj performs between 282 and 380mb/s, cloudflare performs between 186 and 276mb/s. xml: apple 122, ng 181, cf 194, dj 290. x-ray: apple 181, ng 190, cf 209, dj 330. webster: apple 132, ng 195, cf 205, dj 303. sao: apple 234, ng 260, cf 276, dj 380. samba: apple 151, ng 211, cf 221, dj 321. reymont: apple 131, ng 210, cf 213, dj 338. osdb: apple 173, ng 206, cf 224, dj 352. ooffice: apple 142, ng 170, cf 186, dj 291. nci: apple 119, ng 171, cf 191, dj 282. mr: apple 164, ng 187, cf 208, dj 329. mozilla: apple 153, ng 187, cf 199, dj 307. dickens: apple 148, ng 219, cf 238, dj 344.
Travis Downs retweeted
Given:
1. crc32 has throughput 1 on port 1
2. pclmulqdq has throughput 1 on port 5
3. pclmulqdq+pxor can emulate crc32
It seems that fastest crc32 code should divide input in half and issue a crc32 _and_ a pclmulqdq every cycle.
Code and numbers at corsix.org/content/fast-crc3…
Travis Downs retweeted
Filtering numbers quickly with SVE on Amazon Graviton 3 processors lemire.me/blog/2022/06/23/fi…