Stephan Dollberg

🎶 Here by the sea and sand Nothing ever goes as planned Well, in this case our #engineers rolled up their sleeves to make sure things DO go to plan. Redpanda uses #Seastar's heap memory profiler, but it's costly and unusable in production. So, they swooped in with some seriously handy optimizations (that are now merged in the Seastar repo) 💪 The result? Instant detection of what's guzzling the most memory at runtime to swiftly debug and fix OOM issues 🏆 Want to know how they did it? @StephanDollberg tells all here👇 redpanda-data.medium.com/always-on-prod…

@RReverser Right, both are functions but return different types. extract returns foo while std::move returns foo&& (rval reference). Not saying anything in C++ makes sense :) move semantics was just added on in hindsight I guess

@RReverser In take2 you see the destructor call of the copy returned by extract. In take3 the destructor of justdropit. C++ doesn't have destructive moves so move just casts the pointed in object. It only gets destructed later.

@RReverser std::move is just a cast, it doesn't do anything

How Profile-Guided Optimization Makes Your Code Faster Without Any Code Changes - Stephan Dollberg youtu.be/3RtMMHkVsDg #cpponsea #cplusplus #cpp

@dendibakh The ones I referred to above from your posts, i.e.: why can if (foo == bar) baz++ not always be optimized to baz += foo == bar and similar for if (x > y) if (a > b) to if (x > y | a > b) for this I assume you'd need some limit on when it's worth it or not

@StephanDollberg What optimization you mean compilers should do?

Finally, the summary of Performance Challenge #4 is out! Many interesting transformations and >10x speedup of Canny image edge detection algorithm. Check out the 4-part blog post on easyperf blog: #updated-27th-march-2021" target="_blank" rel="nofollow noopener">easyperf.net/blog/2021/02/0… #performance #challenge #perf #cpu #optimization

@Cloudflare What did you use to measure frontend/backend stalls for this? Just what perf stat gave you or anything more advanced like Intel's TMA maybe with input from Altra which counters to use?

ARMs Race: Ampere Altra takes on the AWS Graviton2. cfl.re/3ldX6Hi

Engineering is true magic

Falcon Heavy side cores have landed at SpaceX’s Landing Zones 1 and 2.

Interesting post on llvm-dev about why std::unique_ptr is technically not zero overhead because of ABIs. lists.llvm.org/pipermail/llvm… #llvm

Falcon 9 first stage has landed on Of Course I Still Love You — world’s first reflight of an orbital class rocket.

@lvanbever why is it not on Cloudflare yet :P

@majek04 Nice read, can the thundering herd problem not also be fixed by using REUSEPORT? Wondering how that performs with so many processes

@RReverser @rustlang also compiler bugs

@RReverser @rustlang hidden unsafe code everywhere

list0r web 400 write up dollberg.xyz/ctf/2016/12/29… #33c3 #ctf #33c3ctf #writeup