Orna Agmon <[EMAIL PROTECTED]> writes: > In the advanced OS course in HaifaU Oleg Goldshmidt showed benchmark > results, which show that migration does not enlarge the ability of the > cluster to deal with loads in a significant manner. Back filling, on the > other hand, does.
Ahem... ;-) I would not put it quite so bluntly (and by the way, thanks for listening :). MOSIX is basically a load-balancing infrastructure for clusters. The graphs I showed during the lecture dealt with utilization and average response time et caetera, et caetera, on parallel multicomputers. One can argue that effective load balancing is likely to help with utilization, throughput, response time, and whatnot. However, I cannot state firmly that the results I showed (for a particular workload, in a simulation of a particular machine, using some algorithms involving resource allocation heuristics etc) kill MOSIX. What they do show is that backfilling improves the situation compared to FCFS, and the effect wipes out the differences caused by other potential improvements such as job migration. Now, "migration" in that context is *not* MOSIX-style migration of processes. You have a parallel machine running parallel jobs, and you want to allocate resources for the jobs that sit in your submission queue. One thing that you may consider is checkpointing and stopping some of the running jobs, relocating them to different processors, and restarting, and then allocating the freed processors to new jobs that could not have been accomodated otherwise - think of it as defragmentation or compactification not so dissimilar from what various memory management algorithms do (in OSes, JVMs and elsewhere). Quite different from MOSIX, though one can pull some analogies if one is so inclined. > http://cs.haifa.ac.il/courses/AdvancedOS/ParallelScheduling.pdf > slide 28. I suggest you go to the source for those particular graphs: "Job Scheduling for the BlueGene/L System" Elie Krevat, Jose G. Castanos, and Jose E. Moreira JSSPP 2002 (http://www.cs.huji.ac.il/~feit/parsched/parsched02.html) http://www.cs.huji.ac.il/~feit/parsched/p-02-3.ps.gz to understand the context better. These are just examples I picked from a familiar paper, but the statement that backfilling helps, and helps more than quite a few other tricks, has been made by numerous authors. Again, relevance to MOSIX is tenuous at best. I should apologize for the fact that my slides currently lack proper references, but it's in the works. I hope to add more detailed referencing to the course web page during Hol Ha'moed (well before exam time ;) I will stay away from the whole MOSIX vs. OpenMOSIX issue if you don't mind. ;-) -- Oleg Goldshmidt | [EMAIL PROTECTED] | http://www.goldshmidt.org ================================================================= To unsubscribe, send mail to [EMAIL PROTECTED] with the word "unsubscribe" in the message body, e.g., run the command echo unsubscribe | mail [EMAIL PROTECTED]
