From: <mic...@sy...> - 2002-08-31 15:00:54
|
> I would like to solve that tedious bug > with complete nt database. Did someone > else have my same troubles with it? Yes. There is a conflict between the openMosix blast patch and the pre-formatted nt database. Try downloading nt itself and formatting it with formatdb yourself. Also, I don't have the benchmark numbers handy right now, but my initial tests with and without the openMosix blast patch showed that the patch caused blast to run about 20% slower (single node to single node comparison). So I am not using the blast patch. I went with an MPI solution and am very satisfied with it. Similar to the DFSA solution, I put all the blast databases on every node and then tell the MPI-farmed process to find the blast dataset on its local host. On a related note, when my MPI processors invoke blast, I have been wrapping it in a "mosrun -0", but after seeing the discussion here on using MPI on top of (open)Mosix, maybe I shouldn't? Perhaps openMosix will migrate the MPI processor if there is a less busy node? I'll have to try it... -Michael Peterson -----Original Message----- From: ben To: ope...@li... Sent: 8/31/02 7:57 AM Subject: Re: [openMosix-general] did anyone compare the efficiecy between mosix cluster and standalone pc I would like to solve that tedious bug with complete nt database . Did someone else have my same troubles with it ? I gave a look to openmosixblastpatch diff file and I saw that the patch removes trheads and memory mapped i/o from plain blast to permit dfsa migration.When this is done blast stores its data into buffers rather than into shared memory. As patched blast behaves nicely with smaller databases but fails with the big one can I infer that this might be a buffer lenght / storing / acces problem? I will try to leave mm-i/o support and see what happens .... |