Fermi LAT Gamma-ray pulsar search #4 "FGRP4" App version 1.05

I set up a total of 8 hosts with the v1.05 app. I've accumulated a few days worth of results. I wasn't planning to try to pinpoint the start of the coherent follow-up stage by watching and noting when the final checkpoint was actually written as I expected to be able to see the effect indirectly by averaging a large number of tasks and measuring the "negligible" reduction in overall crunch time. I was planning to assume that all of the gain was coming from a reduced time for the follow-up stage.

The main problem is that what I'm seeing, with one notable exception, is an increase in the overall crunch time and not a reduction. My impression is that the follow-up stage doesn't seem to have changed all that much, except perhaps for Haswell CPUs - see later. Because of the normal variability of this stage anyway, it's impossible to be sure about that without properly measuring it over a significant sample size - a lot more than I've accumulated to this point.

Of the 8 hosts, 4 are new. I took the opportunity to 'run them in' using the beta test app. They will eventually host GPUs, but not yet. They all have overclocked Haswell refresh CPUs (G3258) and seem to be performing very well. The follow-up stage does seem to be quite quick but the whole task is going very quickly anyway.

Of the remaining 4, one is an i3-4130. With the v1.04 app and running all 4 threads, it was taking around 28,870 secs per task on average. The first 20 v1.05 tasks have taken an average of 27250 secs per task - a speedup of over 5%. I'm very much inclined not to read too much into this because you can see these sorts of variations anyway, over the longer term with nothing else changing.

I'm more interested in the remaining 3 hosts, all Intel CPUs but of a much older vintage (4-5 years ago). One is an i3 540 HT enabled, the second is a e6500 Wolfdale dual core, and the third is a Q8400 quad core. Each of those has so far showed an increase in overall crunch time of around 2-3%. Once again, this may well be an artifact caused by the relatively small sample size (~20 results) being averaged.

Each of these three hosts has shown painfully long follow-up stages and I was hoping that the description of "significantly shortened" was really going to help them along. So far, they just seem to have gone backwards :-).

Is there any likely reason why Haswell CPUs seem to be benefiting but the older ones not?