[Disksim-users] Disksim performance

[Greg Ganger]

I can answer a few of the last questions: DiskSim performance tuning
happened some in its very early days (early 90s), with some regretable
long-term code ickiness, and not at all since.  I imagine performance
can be improved by minimizing configurations, such as by minimizing
layers of busses/controllers.  I'm sure that it is far from efficient,
though I was surprised by the 10Kops/second number... doesn't mean that
it's not accurate, but was quite surprising to me.

Greg


On Fri, 3 Apr 2009, John Carter wrote:

> Hi all -
>
>    I am considering extending DiskSim to explore some ideas related
> to improving storage system energy-efficiency using flash.  One
> consideration is DiskSim's simulation performance, since I hope to
> evaluate the ideas on traces with over 1B IOs.
>
>    When I compile and run DiskSim on a 1M-entry trace file
> (ascii.trace replicated 10X with the timestamps updated using the same
> configuration as the runvalid test for ascii.trace), it takes 91
> seconds to execute on 3.2GHz Xeon (quad-core, but I don't think that
> matters since DiskSim is single-threaded, right?).  That translates
> into a bit over 10K simulated IOPs, which is slower than real-time for
> some of the high-end configurations I'd like to explore.  And it
> translates into a little over 24 hours to simulation 1B IOs, which
> could get painful.
>
>    One of my colleagues asked Greg (Ganger) about DiskSim's
> performance at FAST, and Greg apparently replied that 10K simulated
> IOPs was slower than he would have predicted.  He further suggested
> that I contact this list to see if it fits other people's experience.
> What kind of simulation speeds are y'all seeing?  Obviously it depends
> on the configuration being evaluated, but the ascii.parv configuration
> did not look at all onerous.  Are there ways to improve simulation
> performance that a new DiskSim user like myself might not have found?
> For the DiskSim implementors out there, how tightly tuned is DiskSim's
> performance -- has it already been tuned pretty hard (so there isn't
> likely much more performance to squeeze out) or has performance tuning
> not been a focus (so perhaps simulation speed could be improved a lot
> if somebody were to focus on it)?
>
>    Thanks for your input!
>
> Cheers,
> John
> _______________________________________________
> Disksim-users mailing list
> Disksim-users at ece.cmu.edu
> https://sos.ece.cmu.edu/mailman/listinfo/disksim-users
>