3 ms·

Nice article! I wonder if now, with all the NUMA stuff and processors with hundreds of cores something changed sufficiently enough that it warrants another comp

Nice article! I wonder if now, with all the NUMA stuff and processors with hundreds of cores something changed sufficiently enough that it warrants another complete redesign similar to what happened in the article

toast0 1y ago
Much of that is well addressed by one allocator arena per cpu, and either pinning threads to cpus or at least having a high threshold to move threads across NUMA boundaries. If you have a lot of cross thread memory use, maybe you need something to help with allocate on core X, free on core Y and the cross core communication that causes (maybe that's already in place?). There's more memory overhead that way, but large core count systems tend to have a lot of memory too.
- masklinn 1y ago
  > Much of that is well addressed by one allocator arena per cpu, and either pinning threads to cpus or at least having a high threshold to move threads across NUMA boundaries. Note that that can have an awkward effect: if the thread gets parked (either entirely, or just stops calling the allocator because it has reached its steady state), the allocator may never have the opportunity to release that thread's memory. IIRC mimalloc suffers from this issue, you need to call an allocator specific API to tell it about the regime change.
karmakaze 1y ago
A lot of the article talked about swap which wouldn't be a concern in normal operation of most production servers--cache/memory locality still matters but not as dramatically. Back when I was managing bare-metal MySQL servers we were getting scaling to NUMA memory (with jemalloc/tcmalloc). There was an initial performance degradation that required a lot of fine-tuning even working around how the same motherboard/CPUs would initialize core affinities differently. A new problem was deadlocking of large transactions that touched multiple buffer instances. Mind you this wasn't a clean codebase that put a lot of thought into avoiding deadlocks (up until then). At the time I didn't think much about how the allocators could help as they're constrained to the ABI. Writing in Zig with custom allocators for everything would. The only mysql NUMA setting was innodb_numa_interleave=ON which wasn't very good but not a lot worse than trying harder.

toast0 1y ago

Much of that is well addressed by one allocator arena per cpu, and either pinning threads to cpus or at least having a high threshold to move threads across NUMA boundaries. If you have a lot of cross thread memory use, maybe you need something to help with allocate on core X, free on core Y and the cross core communication that causes (maybe that's already in place?). There's more memory overhead that way, but large core count systems tend to have a lot of memory too.

masklinn 1y ago
> Much of that is well addressed by one allocator arena per cpu, and either pinning threads to cpus or at least having a high threshold to move threads across NUMA boundaries. Note that that can have an awkward effect: if the thread gets parked (either entirely, or just stops calling the allocator because it has reached its steady state), the allocator may never have the opportunity to release that thread's memory. IIRC mimalloc suffers from this issue, you need to call an allocator specific API to tell it about the regime change.

karmakaze 1y ago

A lot of the article talked about swap which wouldn't be a concern in normal operation of most production servers--cache/memory locality still matters but not as dramatically. Back when I was managing bare-metal MySQL servers we were getting scaling to NUMA memory (with jemalloc/tcmalloc). There was an initial performance degradation that required a lot of fine-tuning even working around how the same motherboard/CPUs would initialize core affinities differently. A new problem was deadlocking of large transactions that touched multiple buffer instances. Mind you this wasn't a clean codebase that put a lot of thought into avoiding deadlocks (up until then). At the time I didn't think much about how the allocators could help as they're constrained to the ABI. Writing in Zig with custom allocators for everything would. The only mysql NUMA setting was innodb_numa_interleave=ON which wasn't very good but not a lot worse than trying harder.