May 26, 2024

The indifference to RAM and the CPU dependence makes sense, though I went form 32 dual cores to 10, which should have increased the run from 11 mins to 35 (not 27). So we made up for it somewhere else. But it's close and makes sense. However, in none of these 0-300k runs did we get memory bound. Once that happens the number get bad, and that happens pretty quickly at 32GiB RAM above block 300k due to the need to read all blocks and prevouts for validation. In the run from 300-400k with RAM at 512GiB it's 20 minutes. At 160GiB RAM it's 21 mins, basically the same since we don't get memory bound. But in the 400-450k runs with RAM at 512GiB it's 20 minutes, but 60 minutes at 32GiB RAM - with heavy thrashing.