View Single Post
Old 2005-09-26, 17:06   #11
VBCurtis's Avatar
Feb 2005
Riverside, CA

7·683 Posts

When planning to run LLR to a very high value (on a high-weight k, 1M is very high), I find that maximum efficiency is produced by breaking a "chunk" off for LLR when the sieve time is a little less than double the LLR test for that exponent. This is not traditional wisdom, because sieve time per p-value does not scale linearly with range to be sieved-- it scales with the square root of range, so it's more efficient to leave n-values in than it first seems. Since you're planning to run to 1M eventually, I'd sieve 300k-350k until sieve time is quite a bit longer than LLR time-- if LLR is 8 min, I'd go 12 to 14 min on the sieve before breaking those off. This adds sieve time up front, but reduces LLR time at least as much.

Shane-- you say RMA automates this. Does it take this effect into account? Your reply sounds like it breaks off a power when LLR time is equal to sieve time-- this is not as efficient as it could be. Comments?
VBCurtis is offline