search

mratsim / weave

A state-of-the-art multithreading runtime: message-passing based, fast, scalable, ultra-low overhead
Other
532 stars 22 forks source link

Vastly simplify and accelerate memory management. #192

Open mratsim opened 1 year ago

mratsim commented 1 year ago

For a fresh implementation of a threadpool for high-speed cryptography (https://github.com/mratsim/constantine/tree/1dfbb8b/constantine/platforms/threadpool), I found a new design for low overhead memory management:

Assuming we have in a full message passing environment, this is equivalent to sending back the task to the waiter once it is finished hence we are compatible with one of Weave's design goal.

With this, we might be able to completely remove the memory folder and the 2 levels caching structure via memory pool + lookaside list. This might accelerate machine learning algorithm like GEMM / Matrix multiplication as those are very cache sensitive and our memory pool does trigger many page faults.

dumblob commented 1 year ago

This is an interesting idea. Any preliminary measurements showing the difference in behavior in all/many tasks from the weave benchmark set (not just for cryptography computations)?