I have also tested using this loop. For a reasonably large file being read, using the loop executes 300k fewer simulated instructions (~6% of the readfile execution) than using memset. On the other hand, the simulation time of the readfile call using memset was actually 0.4s quicker (~5%) than using the loop. The simulated system still needs to do a pagetable walk and mapping for each page regardless of the implementation. So, my conclusion was that it doesn't make a perceptible difference in performance, and the use of memset abstracts away from possible problems with strided accesses using a static page size.
Joel On Mon, Aug 9, 2010 at 12:50 PM, Ali Saidi <sa...@umich.edu> wrote: > > > > On 2010-08-09 12:13:59, Nathan Binkert wrote: > > > > > How is this not a possible issue for every isa? We're talking about > touching 256kB of data. That shouldn't take very long. We've beaten this to > death, it's time to just call it done and move on. If we're that concerned > just do: > for (int x = 0; x < sizeof(buf); x += 512) > buf[x] = 0; > > It improves the speed by three orders of magnitude, doesn't require and > ifdef and will work on everything that doesn't have a unbelievably small > page size. > > > - Ali > > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > http://reviews.m5sim.org/r/121/#review154 > ----------------------------------------------------------- > > > On 2010-08-09 10:35:49, Joel Hestness wrote: > > > > ----------------------------------------------------------- > > This is an automatically generated e-mail. To reply, visit: > > http://reviews.m5sim.org/r/121/ > > ----------------------------------------------------------- > > > > (Updated 2010-08-09 10:35:49) > > > > > > Review request for Default. > > > > > > Summary > > ------- > > > > util/m5/m5.c: in readfile(), added memset to touch all pages - ensure > they are in the page table > > > > This problem is caused by Linux demand paging. If the pages are not yet > mapped in the page table, the M5 utility does not know the physical memory > address in the simulated system to which it is sending the file read from > the host machine. > > > > > > Diffs > > ----- > > > > util/m5/m5.c a75564db03c3 > > > > Diff: http://reviews.m5sim.org/r/121/diff > > > > > > Testing > > ------- > > > > This fixes the functionality for x86, where the problem was first > encountered. I have also tested the utility for Alpha. The simulated > system executes approximately 10% more instructions during the readfile > operation due to the memset, but the simulation time required for this is > still marginal. Using memset provides an ISA independent solution compared > to buffer accesses that use a page-sized stride. > > > > > > Thanks, > > > > Joel > > > > > -- Joel Hestness PhD Student, Computer Architecture Dept. of Computer Science, University of Texas - Austin http://www.cs.utexas.edu/~hestness
_______________________________________________ m5-dev mailing list m5-dev@m5sim.org http://m5sim.org/mailman/listinfo/m5-dev
