Jeff Plaisance

@geofflangdale Yep, and the "interpreter-like" approach is generally what is done, but in many circumstances it leaves a lot of performance on the table over human-written SIMD code for the specific query. RTCG has its own issues though, like balancing compilation time against execution time.

@jeffplaisance to always find primitives that are big enough that composing them via "interpreter-like" mechanisms isn't a particularly big penalty. But I can see that for truly optimal performance on "not known in advance" queries it might be necessary to take the RTCG path.

Thinking a bit about "pipeline design" for software. If we've got a reasonably complex problem, we might have a pipeline with 4-5 major stages and perhaps dozens of instructions. I spend a lot of time struggling with not just how the problem should be solved, but how it ...

@geofflangdale On a related note, do you happen to know why _mm512_mask_compressstoreu_epi8 is way slower than _mm512_maskz_compress_epi8 followed by _mm512_mask_storeu_epi8 on zen 4 (and possibly other microarchitectures?)

without being mugged by some ludicrously slow operation. Most vendors seem to have perpetrated this at some stage: it's not a dig at any particular company.

Radical idea: ISAs come with performance bounds. Implementations outside the bounds (which might be expressed in terms of 'relative cost' of operations: e.g. this double-shuffle is no more than 6x the latency and reciprocal throughput of an ADD or AND) are non-conforming.

@EmilStenstrom @lemire @Love2Code the "hot path" is a myth. you should read this post, written by the guy you are replying to: lemire.me/blog/2023/04/2…

@lemire @Love2Code To add some more nuance: compute is only relevant when the code you run are in the “hot path”. For almost all _web_ applications, that’s database queries and network requests. The backend language is mostly just waiting.

Reddit: "accessing our services is extremely costly to us, so here is a bill for 20M$" arstechnica.com/gadgets/2023/0… Me: "I wonder what the backend of reddit is like..." support.reddithelp.com/hc/en-us/artic… Python, eh?

I've open sourced my B++ Trees library, which is a B+ tree library that I wrote in C++. It can be used as a normal B+ tree, but it also has optional mixins that can be used to access elements by index, calculate prefix sums, or find the min/max element github.com/jeffplaisance/…

@xeraa @fulmicoton @papers_we_love there is a third option of using userfaultfd and madvise to get the benefits of hardware accelerated virtual address resolution with more control over eviction than is possible with mmap. still has some of the downsides of both but could be the right tradeoff for many systems.

interesting take on MMAP. so is the counter take ("compares apples to camels"): ravendb.net/articles/re-ar… PS: looking forward to when we can do @papers_we_love vienna meetups again (in-person only)

@GrahamJenson @ohunt if you squint at it hard enough it's just insertion sort with some extra unnecessary steps. it still works if the inner loop goes from 0..i although the intermediate steps turn out a bit different.

x = [5,2,4,3,1] for i in range(len(x)): for j in range(len(x)): if (x[i] < x[j]): x[i], x[j] = x[j], x[i] # swap print(x) # [1, 2, 3, 4, 5] A weird sorting algorithm from arxiv.org/pdf/2110.01111… via @ohunt How this works?

@seldo @DavidKrakov vega-lite (vega.github.io/vega-lite/) is kind of this for individual visualizations, although to my knowledge it's not widely adopted across tools

@DavidKrakov Fascinating idea!

What I've learned in the last year about how data teams should work: seldo.com/posts/what-i-v…

@fulmicoton not sure but the best source for this type of info is the intel optimization manual

Super newbie branch prediction question. Can a CPU engage in more branches when it is doing speculative execution? If yes, is there a depth limit?

@DGWorldTour @JussiMeresmaa Thanks so much for resolving the signup issue for the international players at the European Amateur Open!

@JussiMeresmaa Annoyed about European Amateur Open signup. Site said it opened at 3 pm CEST but it was full by 2:45. Wrong timezone on site.

@DGWorldTour Annoyed about European Amateur Open signup. Site said it opened at 3 pm CEST but it was full by 2:45. You put wrong timezone.

Planning your exit? Job searches on @indeed to Canada were up 10x in the hours after election was called.

@jeffplaisance et al present a paper on a faster decoder for VByte-compressed data. Read more: engineering.indeed.com/blog/2015/03/v… #CompSci

@jedisct1 Revised answer: HDFS support isn't done in the open source version yet. This file github.com/indeedeng/imho… needs to be filled in.

@jedisct1 Yes the docs only cover AWS setup right now but we will be adding instructions for non-AWS installations in the future.

@jedisct1 Yes it can, it can use HDFS as an alternative to S3 for shard storage.

Proud to annouce that Indeed has open sourced our distributed analytics and machine learning platform, Imhotep. go.indeed.com/uvque

@doug_gray @naelyn @IndeedEng SSSE3 is my favorite instruction set but AVX2 is a close second

my twitter got haxed... some of that stuff was kind of offensive but it wasn't me :(