Deadline oddity on O1AS20-100T work
20 Feb 2016 15:28:51 UTC
Topic 198448
(moderation:
Ever since the welcome extension of the deadlines for newly sent O1AS20-100T a few days ago, I've seen an oddity that some WUs arrive with a deadline 7 days after send date, and some with a deadline 5 days after send date. Arguably this mismatch may be intentional for some purpose, and thus not properly either a problem or a bug, but it does seem odd.
In my case, I have four hosts, and it seems that three of them have consistently received seven day deadline work in recent days, and one has received nothing but five day deadline work. This distribution by host may be entirely accidental, but seems a bit of a long coincidence.
Similarities:
1. all four hosts are Windows, three being Windows 7 and one Windows 10.
2. all four hosts run BOINC 7.6.22
3. all four hosts have run only Einstein as a BOINC project for weeks, months, or ever.
Differences:
1. The 5-day deadline host is a Westmere CPU, thus not AVX capable, and has been receiving SSE2 work, while the other three are differing flavors of more modern base Intel CPU types, and have all been receiving AVX since a few hours after V1.03 work issue started.
2. Three of the four have a pair of GPUs each, while one is a laptop without BPU, but the 5 to 7 split does not align with that.
In case this is of any interest, I'll include a link to the current O1AS20-100T task list for each host:
7-day host1 aka Stoll7 (four-core Sandy Bridge)
7-day host2 aka Stoll8 (two-core Haswell)
7-day host3 aka Acer2 (thelaptop) (two-core Sandy Bridge)
3-day host1 aka Stoll6 (four-core Westmere EP)
Extra note not relevant to the posed issue: The irregular fetch pattern and considerable number of aborts are artifacts of my decision changes on participation (I had been running zero CPU jobs for power efficiency reasons) and difficulties of over-fetch (I only allow a single CPU job, and the system seems systematically to over-fetch in my case, possibly because some BOINC component does not adjust down my presumed capacity to reflect that, which is especially a problem with the shortish deadlines on the O1AS20-100T work)
Deadline oddity on O1AS20-100T work
)
Workunits generated earlier have a 5d deadline (initially even 3d), workunits generated later (after - IIRC - Feb 15) should have 7d. There are still tasks being produced from early workunits, when previously sent tasks are reported as errors. Locality scheduling leads to similar tasks ending up on the same host.
BM
BM
RE: Workunits generated
)
Second false alarm from me due to failing to think of the locality scheduling effect--sorry. As you suggest, a review of the recent work units from the "5-day" host shows they were created very near 15 Feb 2016, 17:15:00 UTC.
By contrast a review of creation dates for the WUs recently received by the 7-day hosts shows them created on 16 Feb 2016, between 14 and 15 UTC. On a given host I see extreme creation date clustering, suggesting those WUs were part of a locality grouping.
Thanks for guiding me away from my confusion.