im working in smp. On Sep 16, 2011, at 8:46 AM, erik quanstrom <[email protected]> wrote:
> hey, ron. > > On Fri Sep 16 01:57:04 EDT 2011, [email protected] wrote: >> for the 2M pages -- I'm willing to see some measurement but let's get >> the #s -- I've done some simple measurements and it's not the hit one >> would expect. These new machines have about 10 GB/s bandwidth (well, >> the ones we are targeting do) and that translates to sub-millisecond >> times to zero a 2M page. Further, the text page is in the image cache. >> So after first exec of a program, the only text issue is locating the >> page. It's not simply a case of having to write 6M each time you exec. > > however, neither the stack nor the heap are. that's 4MB that need to be > cleared. that sounds like an operation that could take on the order of > ms, and well-worth measuring. > > maybe it might make sense to use 4k pages for stack, and sometime for > the heap. > >> I note that starting a proc, allocating and zeroing 2 GiB, takes >> *less* time with 2M pages than 4K pages -- this was measured in May >> when we still were supporting 4K pages -- the page faults are far more > > i'll try to get some numbers soon, but i think a higher priority is to get > a smp setup. is anyone testing with smp right now? > > are you sure this isn't the difference between throughput and latency? > did you try a small-executable test like > for(i in `{seq 1 1000000})dd -if /dev/zero -of /dev/null -quiet 1 > ? now that the code has been removed, it's a little harder to replicate your > numbers. > >> expensive than the time to write the memory. Again, YMMV, esp. on an >> Atom, but the cost of taking (say) 6 page faults for a 24k text >> segment that's already in memory may not be what you want. >> >> There are plenty of games to be played to reduce the cost of zero >> filled pages but at least from what I measured the 2M pages are not a >> real hit. > > i'm okay with the atom suffering a little bit (odd how far down the food > chain one can get 64 bits!), i'm actually more concerned about > being punshed severely for forking on a beefy but busy machine. > the atom is just my test mule. > > (can't one just preemptively map the whole text on first-fault?) > > - erik > > p.s. i guess the claim i thought i saw that you need 1gb pages > isn't correct. that's good, but i need to track down why i don't see 'em on > the atom. >
