May 29, 2026

In terms of what we expected, the idea is not necessarily less computation but less RAM (non-mmap) utilization during download for a given number of blocks. The idea was that this frees up mmap for table heads, which need to remain mostly RAM resident) and working set for paging to support indexation and validation. Given much higher RAM I didn't expect this to have a material impact on the maxi, more on the mini and pi. With more available mmap paging is reduced and we can avoid thrashing at a lower RAM level. At the same time we reset our benchmark baseline to 950,000 blocks. The pi is @eynhaender’s new toy, with just 16GB RAM and 4 slow threads (~2.1GHz, same as the maxi, but with sha-ni, unlike maxi).