Oct 19, 2026

This has significantly improved low end performance. macOS pruned 10GB milestone dropped from 72 mins to 41 mins. The linux/windows milestones didn't change, as they remain network limited. Validation time on the minimum RAM (18GB) mac dropped at least 2 hours (partial run). So I'm going to drop the RAM schedule for macOS mins from 10/18GN to 8/16GB to match the others, and then reverify. This will result in consistent default head sizings by RAM on any OS. Also, all of the bigmac runs in the past, and these new ones, were running with AVX-512 enabled. But bigmac has the crippled chip, and no shani, so this materially hurts performance. The new startup config code highlighted the issue. So the verification runs will be compiled with AVX512 off (AVX2), and will certainly benefit from that as well. Finally, the strong table (head only) is written in purely sequential order, and read in the validation partial order (organized by concurrency window), but was still configured as random access (hashmap). So this will also be updated for the verification runs and should assist validation further. Total store size shrinks by 11.7GB, but the head gets almost 2GB larger. This slows snapshot/recovery a bit (maximal head set at 950k now 19.9GB). But the head being an arraymap means that its just streaming appends to the head as blocks arrive, total write ordering. So we don't force this head to be resident. That change was verified to be very effective. It is flushed to disk quickly after write. So the net head-resident space actually shrinks.