Conversation

I honestly wonder if there are changes to CPU and memory hierarchy design that could make GC less of a sin.

Quote Tweet

Dec 16, 2021

"[We find] situations where low-pause GCs ... achieve worse application latency [and much worse throughput] than simple stop-the-world GCs" arxiv.org/pdf/2112.07880

Replying to

I think it's more about language design and OS APIs than hardware support. userfaultfd and MREMAP_DONTUNMAP haven't been around for long and help a lot with compacting GC without requiring new hardware. MADV_FREE is useful too. Hardware has a lot of capabilities not being used.

Replying to

and

Languages can be designed around thread-local heaps with special types for either transferring ownership of data to another thread or sharing data. JavaScript is essentially designed that way with the ability to transfer an ArrayBuffer or to explicitly use a SharedArrayBuffer.

3:52 PM · Dec 16, 2021Twitter Web App

Replying to

and

There's much less overhead if nearly everything is on fully separate thread-local heaps unable to reference each other. So much of the overhead comes from sharing a huge heap between threads. Short-lived tasks would be able to completely avoid ever actually collecting garbage.