Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Four Kinds of Optimisation (2023)

tratt.net

Post reply on HN
Print view View original

11–16 of 16 posts

Re: Four Kinds of Optimisation (2023)

#12

Post by Supermancho »

genman wrote: post #8
sirwhinesalot wrote: post #6

This is missing the 5th and most important optimization technique: don't pessimize the code in the first place. The usual culprit is "premature modularization", where code that is used in one place and is never going to be extended is nonetheless full of indirections.

In principle it can in general fit the points 1-3 when you view less abstractions as lower level system and code as a data structure and algorithm, what can also include different levels of parallelism.

> In principle it can in general fit the points 1-3

In principle, I don't think people would lump it in.

Re: Four Kinds of Optimisation (2023)

#14

Post by pyfon »

willvarfar wrote: post #4

I've spent a lot of time jumping in to optimise horrendously complicated programs and data pipelines and things. The big wins are from understanding the domain and spotting all the inevitable misunderstandings and unnecessary steps that creep in when systems are so big they get split up and built/maintained by different people at different times.

This! Extending the list of 4:

5. Remove unrequired behaviour.

6. Negotiate the required behaviour with stakeholders.

7. UX changes. E.g. make a synchronous flow a background job and notification. Bring back quick parts of the operation sooner (e.g. like a progressive jpg)

8. Architecture changes. E.g. monolithification, microservification. Lambdas vs. VM vs. Fargate etc.

And some more technical:

9. Caches?

10. Scalability, more VMs

11. Move compute local (on client, on edge comoute, nearby region)

11a. Same as 11 for data residency.

12. Data store optimisation e.g. indices, query plans, foreign keys, consistent hashing (arguably a repeat of data structures)

13. Use a data centre for more bang/buck than cloud.

14. Compute type: use GPU instead of CPU etc. I'll bundle here L1 cache etc.

15. Look at sources of latency. proxies, sidecars, network hops (and their location) etc.

16. GC pauses, event loops, threading, other processed etc.

Re: Four Kinds of Optimisation (2023)

#15

Post by bytepoet »

The blog post is really good. I see that there's a follow-up piece 'The Fifth Kind of Optimisation' about parallelism.

Something that I'd like to add is that it's helpful to understand the optimization capabilities of our compiler. Ideally, we would like to write program that doesn't have what are called 'optimization-blockers' - these make it hard for the compiler to generate optimized executables.

I like the pointer to the blog on accidentally quadratic implementations. I find that the following pattern is often a landmine:

for (int i = 0; i strlen(s) gets computed every iteration, incurring O(n) time.

Finally, being aware that I/O latencies are major source of bottlenecks leads to nice optimizations. One advantage of multiple threads is that they can sometimes hide the I/O latency. In general, writing programs with good memory locality is one of the better levers for optimization.

Post reply on HN