What's next?
Anonymous
8 Aug 2007 7:39:33 UTC
Topic 13579
(moderation:
A new run, S5R3, will follow S5R2.
We're still preparing this and not all decisions are final yet, but the current state is this:
- It will use the same data set as S5R1 from S5 up to Jan 2007. In principle more data gives us more sensitivity, but we decided to wait for the end of S5 so we can pre-process and analyze all of the S5 data (a full year netto!).
- It will cover a different, wider range of frequencies than S5R2 (up to about 800 Hz)
- Instead of juggling with tiny frequency ranges as we do now we will probably also split up the sky between workunits. This may give a somewhat odd movement of the direction marker in the screensaver graphics, but will gives us better control of computing overhead an the length of the workunits. It will even make the post-processing rather easier than harder.
S5R4 is already being thought on. It will probably cover the whole of S5 data, and should then (again) be the most sensitive search ever performed for gravitational waves.
It seems the cross-platform differences (observable as validation problems) are sorted out and fixed now, but apparently it takes much longer than we expected to find and fix the remaining problems in the new code. As long as these problems persist I'm afraid to introduce even more by adding more platforms and speedup measurements to the existing code, so "optimized" Apps are unlikely to come up before S5R3.
BM
What's next?
)
I doubt that they will come in the remaining days of S5R2, but will hopefully come early during S5R3.
BM
BM
Some update on S5R3: - The
)
Some update on S5R3:
- The run-time of the workunits will be roughly the same we had at the beginning of S5R2 (12h at max on our reference machine - sorry, don't remember the credit values from the top of my head - was it 190?). We hope to reduce it even further when we put more work in optimizing the Apps.
- The amount of data necessary for a workunit will raise to 35-55 MB (depending on the base frequency). For the lower frequency bands we will use the same data files than we did for S5R2, so we intend to switch off deletion of the S5R2 data files on your machines at some point and you don't have to download a whole completely new set of data files again, just the ones you don't already have left over from S5R2.
- We also think about modifying the scheduler so that the workunits that require smaller amount of data are given to machines with slower connections. However given our previous experiences with scheduler hacks we might not be able to do this at all or on time, given that the scheduler is a mission-critical component that can easily bring down e.g. the database.
BM
BM
The plan is that at some time
)
The plan is that at some time at the end of this week we'll force the S5R2 workunits generator to write all remaining S5R2 workunits into the database (which will make it stop), and then start the workunit generator for S5R3.
The effect of this would be that for some days there will be both S5R2 and S5R3 workunits in the database of which no task has even been sent to a host yet. The transition will be rather smooth and long.
BM
BM
RE: The amount of data
)
Done.
We've generated the remaining S5R2 workunits and put them into the database. You'll find the S5R2 work generator isn't running anymore (Server Status page). This allows us to start the Work generator for S5R3 soon.
BM
BM