muke

computer architecture phd student

vroom vroom

[muke @muke1010101]

Pleased to share my first ever paper :) 'Improving Memory Dependence Prediction with Static Analysis': https://t.co/j7hpnnlb43

[muke @muke1010101]

This is a really great + beginner friendly article on the topic for anyone interested:
https://t.co/3sedMTnFOZ

[muke @muke1010101]

the next exciting thing im working on now is gonna be multi-branch prediction, i.e. predicting more than 1 branch per cycle to keep up with the fetch bandwidth demands of next generation CPUs. This has been announced to be included in Zen 5 too, so it's a hot topic atm

[muke @muke1010101]

on the bright side the feedback from the more positive reviews was useful, so i'll include that for a submission to HPCA in a couple months. hopefully at some point i can just be done with this project entirely

[muke @muke1010101]

tragically this didn't go well, got 2 strong rejects which was enough to not have the option for rebuttal. Frustratingly, it felt as if the reviewers giving strong rejects didn't understand that in doing so they were preventing me from addressing their very simple questions.

[muke @muke1010101]

meant to be getting the reviews for my MICRO submission today😵‍💫

[muke @muke1010101]

same anxiety as waiting for exam results. its not the end of the world if i don't get in but i'm not looking forward to having something i obsessed over torn apart.

[muke @muke1010101]

meant to be getting the reviews for my MICRO submission today😵‍💫

[muke @muke1010101]

Also, the icache prefetcher should really be fetch-directed prefetching with a decoupled frontend, which Gem5 does now have but its kind of broken and ruins BTB accuracy for now. But if that ever changes definitely include that too.

[muke @muke1010101]

Table of Gem5 params for a paper I'm writing. The large model is meant to be a modern high-end workstation/server core. Hoping this helps lay out all the many small things you wanna be tuning in Gem5.

[muke @muke1010101]

Not listed here: the default L2 latency is 40 cycles, which is what the L3 should be (!), set that to like 14. Also the L1-I can get away with 1 because in reality you'd use a u-op cache. You *can* use DDR5 but DDR4 actually has lower latency, which is better for spec.

[muke @muke1010101]

Waiting for work by someone in my research group to get published and then it'll also include the state of the art in data prefetching too

[muke @muke1010101]

This has now been updated to be based off Gem5 v25 which includes a decoupled frontend. I'm having a lot of fun making this better and better tbh.

[muke @muke1010101]

Been tuning a bunch of things in Gem5 for a paper lately and just pushed+listed them all in my fork, includes things like having an actual L3 cache

[muke @muke1010101]

just so many more graphs i need to plot in all my future works lol

[muke @muke1010101]

wow, why did no one tell me gem5 v25 implemented a new store sets

[muke @muke1010101]

ok somebody made it have tagged associative entries now, this is breaking my brain a little

[muke @muke1010101]

I actually think using LLVM isn't fair here because flang is so bad of a frontend that it leaves so many opportunities for PGO to pick up. But just looking at the C/++ workloads it seems to almost always make it worse!

[muke @muke1010101]

Been playing about with PGO on spec2017 and this is what Gem5 says it does to performance:

(profiles came from the train inputs and compiled with LLVM 22)

[muke @muke1010101]

tempted to say i wish it came out from the start of my phd but, maybe for the best i was forced to learn how to do this stuff decently by hand before having it automated

anyway, what we really need now are agents that are integrated into your environment

[muke @muke1010101]

i have to say phd work has become twice as easy since claude, all the tedious python scripting for wrangling results and plotting graphs are like the perfect use case

[muke @muke1010101]

because a 10x'er has never used either 😌

[Joel 🇦🇺 @ptr_to_joel]

you can measure how good an engineer is by their opinion on the jvm / java

[muke @muke1010101]

On the other hand they removed x264 which was disproportionately my biggest performance win. They replaced it with flac but as that works on 1D data instead of 2D there'll probably be fewer independent loads for my work to win on </3

[muke @muke1010101]

Spec2026 is released. I look forward to figuring out all the new miscellaneous problems like 'bwaves needs 16GB of stack space' and 'xz compiles to insane bitshifts that break gem5's X86 frontend'

[muke @muke1010101]

Seriously though some of the new workload selections look cool. I'm surprised it took this long to switch from perl to python, but it makes way more sense + it'll be nice not to have to disable strict aliasing and LTO. It also includes Gem5! So I'll be simulating Gem5 on Gem5