(a) make DB queries more efficient using name>'FILE__' and name<'FILE__~' rather than
name like 'FILE__%'
(b) Set 'no remaining work for this file' flag correctly by making a DB scan if needed.
One can show that this is the 'cheapest' reliable place to put this scan.
(c) Modify deterministic algorithm for finding unsent results so that instead of
starting with FILE="" and scanning forward over all files, it starts at a random
place in file space, scans cyclicly to the end, and then from "" to the start
point.
(d) Satisfy work request if possible. Don't terminate sending work until none left that
is feasible, or request satisfed.
(e) If a new file is needed, first pick file associated with unsent results which are more
than 2 hours old. Note: need to make this a user-configurable option, and add some
random +- slack.
For the record, here is the current locality scheduler logic.
I will update the docs once this is a bit better tested and
stable.
(1) If there is an (one) unsent result which is older than
(1) config.locality_scheduling_send_timeout (7 days) and is
(1) feasible for the host, sent it.
(2) If we did send a result in the previous step, then send any
(2) additional results that are feasible for the same input file.
(3) If additional results are needed, step through input files on
(3) the host. For each, if there are results that are feasible for
(3) the host, send them. If there are no results that are feasible
(3) for the host, delete the input file from the host.
(4) If additional results are needed, and there is (one) unsent
(4) result which is older than 2 hours and is feasible for the
(4) host, send it.
(5) If we did send a result in the previous step, then send any
(5) additional results that are feasible for the same input file.
(6) If additional results are needed, select an input file name at
(6) random from the current input file working set advertised by
(6) the WU generator. If there are results for this input file
(6) that are feasible for this host, send them.
(7) If additional results are needed, carry out an expensive,
(7) deterministic search for ANY results that are feasible for the
(7) host. This search starts from a random filename advertised by
(7) the WU generator, but continues cyclicly to cover ALL results
(7) for ALL files. If a feasible result is found, send it. Then
(7) send any additional results that use the same input file. If
(7) there are no feasible results for the host, we are finished:
(7) exit.
(8) If addtional results are needed, return to step 4 above.
svn path=/trunk/boinc/; revision=5129
actually not impossible. Consider the following scenario: WU A
has result 1 and WU B has result 2. These are both sent to a
host. Some time later, result 1 fails and the transitioner
creates a new result, result 3 for WU A. Then the host requests
a new result. The maximum result already sent to the host is 2.
The next unsent result (sorted by ID) is #3. But since it is
for WU A, and since the host has already gotten a result for WU
A, it's infeasible. So I think this is only wacky if
!one_wu_per_result_per_host.
- David, I simplified the inner part of send_results_for_file()
somewhat. I can't see the need/use for the bool bool
in_working_set argument. If I have really screwed the pooch
please revert.
svn path=/trunk/boinc/; revision=5106
- bug fix 1, trivial, n % 0 generates FPE if no files.
- bug fix 2, sublte. In the deterministic hunt to find
the next unsent result, the comparison of the result name must be
done with the lexically maximum result name constructed from the
given filename. Eg, if one needs name>FILE_ZZZZ...Z where
Z=0xff (is this charset independent??) and the total name length
is 255 (or is it 254??).
- David, I'll clean out some of the debugging scaffolding in a few
more days, when I get some confidence that there are not other
problems to fix! Please read embedded comments/questions in code.
svn path=/trunk/boinc/; revision=5086
Trivial bug, FPE on n % 0 when host has no files.
Hard bug, in the deterministic search to find a new result that can
be sent, the upwards search on name must be done not by comparing
RESULT name to FILENAME, but instead by comparing result name to the
maximal lexical resultname that can be constructed from the
filename, which is filename_ZZZ...Z where Z==0xff.
svn path=/trunk/boinc/; revision=5085
flag a file as over unless the WU generator has already indicated that
no further work can be remaining. Search code for 'David' to find some
comments.
svn path=/trunk/boinc/; revision=5077
Rom Walton