Hi All, Although i didn't expect any differences, i ran the test in a 64bit kernel with 32bit userspace. It turned out that test 2 and 3 show the difference nicely.
Test 2 is kernel based, and although the difference is small, all results are higher than 32 bit. Test 3 is user-space based, and here you see quite similar results to the 32 bit test, which i think is because user space is still 32 bit. So far in my opinion 64bit SMP is the way to go. The results PDF is attached to this email. Best regards, Wesley --- On Sun, 6/16/13, Wesley PA4WDH <pa4...@yahoo.com> wrote: > From: Wesley PA4WDH <pa4...@yahoo.com> > Subject: Re: [Soekris] net6501 : CPU architecture ? > To: soekris-tech@lists.soekris.com > Date: Sunday, June 16, 2013, 11:37 AM > Hi All, > > @David: I hope you don't mind the reply-to-list. > > Thanks for your suggestion. I agree that packet handling > would be an interesting topic, but i can't really get an > idea on how to actually test it. Of course, suggestions are > welcome :-) > > In the meantime a made a script and did some general > benchmarking. > 1) dd'ing from /dev/zero to /dev/null, to see if and when we > reach full memory bandwidth > 2) dd'ing from /dev/urandom to /dev/null, to see how a > kernel space CPU intensive task differs between SMP and > non-SMP > 3) Calculating 4000 digits of PI using BC, to see how a user > space CPU intensive task differs between SMP and non-SMP > > All tests are done three times: With 1 proces, with 2 > concurrent processes and with 3 concurrent processes. Of > course these processes are doing the same thing, they are > not 3 threads working on the same process. So in case of > test 3 we're calculating PI 3 times. > > The script i made to do this is attached to this email, also > a PDF export of the results is attached to this email. > > *** warning for the script *** > I used "killall" to kill processes, on Linux this allows you > to kill processes based on their name, but i understood that > within *BSD it does what it says on the box and kills all > processes on the system. > > My conlusions: > Based on test 1 you can see a pure memory bound process > doesn't gain or loose much with SMP. However, it never > reached the full 1500 MB/s reported by memtest86+. > Based on test 2 and 3 you can see a clear difference between > SMP and non-SMP when it comes to CPU bound tasks. > > However, when i relate this to your specific case: > When there was only one process running (ie. the packet > handling process) there is barely a difference between SMP > and non-SMP. However, i think you may gain from SMP when > there are other processes running at the same time (syslog, > whatever). > Since there's not much to loose i'd go for SMP anyway. > > I don't think this will reflect 32bit vs. 64bit nicely, > suggestions this this are welcome. > > Best regards, > Wesley > > > --- On Sat, 6/15/13, David Ruggiero <thatseattle...@gmail.com> > wrote: > > > From: David Ruggiero <thatseattle...@gmail.com> > > Subject: Re: [Soekris] net6501 : CPU architecture ? > > To: "Wesley PA4WDH" <pa4...@yahoo.com> > > Date: Saturday, June 15, 2013, 11:19 PM > > Wesley - > > > > I'd be interested in any single-threaded benchmarks > with SMP > > vs > > non-SMP - especially those involving ethernet packet > > handling. > > "Single-threaded" because I'm curious if SMP will still > help > > by > > allowing the OS system functions to use some of the > other > > Hyper-thread > > core, and if that happens if it's enough to overcome > the > > native extra > > work that dealing with SMP causes. "Packet handling" > because > > the > > overhead of interrupt service is something I'd like to > > include in > > that. > > > > You could do some sort of fake loopback using a > crossover > > cable coming > > out of one eth port back into another. > > > > Though I'm using openBSD, this would be applicable to > those > > of us > > using our machines mostly for PF (which is > > single-threaded). > > > > But really, anything you come up with will be of > interest! > > > > cheers > > david > > > > > > On Fri, Jun 14, 2013 at 10:43 AM, Wesley PA4WDH <pa4...@yahoo.com> > > wrote: > > > Hi All, > > > > > > I'm sorry to pick such an old topic again. > > > > > > --- On Wed, 5/15/13, David Ruggiero <thatseattle...@gmail.com> > > wrote: > > >> A few months ago I asked the question here > about > > what the > > >> pros and > > >> cons would be of running the OpenBSD SMP > kernel > > (bsd.mp) on > > >> the > > >> net6501, versus the standard uniprocessor > kernel - > > seeing > > >> that the > > >> E6XX series has HyperTransport and therefore > > something > > >> approximating > > >> multiple cores. There wasn't, as I recall, > any > > consensus, or > > >> even much > > >> information at all (*). So from this > discussion, as > > far as I > > >> can tell, > > >> I have a choice of _four_ different OpenBSD > kernels > > that > > >> could > > >> legitimately be booted on my net6501 and > would > > probably > > >> run: > > >> > > >> 32-bit i386 uniprocessor > > >> 32-bit i386 SMP > > >> 64-bit amd64 uniprocessor > > >> 64-bit amd64 SMP > > >> > > >> I'm sure a similar list of possible > alternatives > > exists for > > >> other BSD > > >> variants and for Linux. > > >> > > >> A different company than Soekris might do > some > > quick > > >> testing, or at > > >> least provide some pithy engineering insight > and > > give its > > >> customers a > > >> quick rundown from their point of view on the > > advantages > > >> and > > >> disadvantages of 32 vs 64bit and non-smp vs > smp - > > whether in > > >> general, > > >> or for specific common application needs. But > at > > the risk of > > >> being > > >> flamed here, I'll say that I'm not holding my > > breath for > > >> any > > >> information like that from the company > anytime > > soon... > > >> > > >> If anyone has data points around these > questions > > from actual > > >> personal > > >> experience and actual real-world testing, > like > > Kyle's, I'm > > >> sure many > > >> would be grateful for that. > > > > > > I just received my 6501 and i don't have any disks > yet. > > However, i'm able to boot into 32bit 64bit linux > kernels and > > i'm able to enable/disable SMP. > > > The kernel version will be 2.6.37 with some > Gentoo > > patches. > > > > > > Since i can't yet use it for what i planned to do > with > > it i'm happy to do some performance testing. Anything > that > > doesn't take too much disk space or requires > installation of > > additional packages is possible. Just keep in mind > that > > whatever i boot, userspace will always be 32bit. If > it's > > really nescessary i can see if i can change that but i > > rather not. > > > > > > Any test suggestions ? > > > > > > Best regards, > > > Wesley > > > _______________________________________________ > > > Soekris-tech mailing list > > > Soekris-tech@lists.soekris.com > > > http://lists.soekris.com/mailman/listinfo/soekris-tech > > > -----Inline Attachment Follows----- > > _______________________________________________ > Soekris-tech mailing list > Soekris-tech@lists.soekris.com > http://lists.soekris.com/mailman/listinfo/soekris-tech >
results.pdf
Description: application/download
_______________________________________________ Soekris-tech mailing list Soekris-tech@lists.soekris.com http://lists.soekris.com/mailman/listinfo/soekris-tech
