Nov 21, 2025

Back up and running on p2p, with the generated dispatch. Performance is the same, as expected. Interestingly I ran it first without the custom memory allocator, which I forgot to reconnect. It was running at about 150% of the normal time. I hadn't looked at the impact of the allocator for a couple years. When I wrote it the impact was not as significant, because there were other issues. But as we've leaned out it's become more impactful, significant in fact. It was written originally to allow blocks to be move to a common strand, which pushed destruction to another thread. That led to huge deallocation costs. We abandoned that approach, but kept the allocator (which had resolved that issue), because it provided a modest perf benefit. Now that benefit a reduction in sync time by about 1/3. I can see this in the flame graph as well, it was at about 32% total time, just for block deallocation. Now it's back to about 7%.