Hi Akira, Thanks a lot for the detailed information! I will try to deploy the verification tools (litmus and herd) next week and let you know if they work. The PPC server has 8 cores and is currently running Ubuntu 16.04, which I'm familiar with. Would that be OK? Or do you have any other suggestions? BTW, I have successfully applied an ARMv8 server with 96 cores from packet.com. Since I'm very interested in how non-blocking algorithms perform on real hardware, I will try to deploy the sample code in perfbook on all three platforms to see how they work. One interesting thing I noticed after a quick check is that, for ARMv8, primitive FAA (__atomic_fetch_add) is performed by using a retry loop that employs LL/SC, and hence it could lead to a performance degradation (roughly, 10x slower when 30 cores perform FAA concurrently). I will post these finds in the list later :-) Thanks, --Junchang On Sat, Oct 6, 2018 at 7:47 AM Akira Yokosawa <akiyks@xxxxxxxxx> wrote: > > On 2018/10/04 08:32:23 -0700, Paul E. McKenney wrote: > > Hello, Junchang, > > > > Glad it worked out for you! > > > > Note that the DIY toolset (http://diy.inria.fr/) has things that > > automatically exercise litmus tests on real hardware. See the > > CodeSamples/formal directory in the perfbook archive, and especially > > its herd and litmus subdirectories, for some example uses. > > (And kudos to Akira for setting this up!) > > Hi Junchang, > > At the time I worked in those subdirectories, "ocaml" 4.02.3 for > powerpc was not capable of natively building DIY toolset. > > Current version of ocaml can build the toolset, which means > you can test litmus tests without cross compiling on powerpc. > > But the default target of Makefile under the "litmus" subdirectory > assumes x86 architecture. For powerpc, you need to use the target > "cross-pcc" even if you are on a powerpc platform. > > One thing I'd like to advise you is that klitmus7 can generate > kernel module sources which can deadlock. Such deadlock can be > hard to escape on a remote virtual machine. > > herd7 doesn't detect deadlock condition. litmus7 runs litmus tests > in userland, and it is easier to kill if deadlock happens. litmus7 > is much slower than klitmus7, but it can help in identifying > possible deadlocks. > > Or by running a particular litmus test by klitmus7 on a host > which can be easily reset if necessary, you can see if it can > deadlock. > > NOTE: api.h in the subdirectory "litmus" predates the merge of > tools/memoroy-model into Linux kernel, and covers limited subset > of kernel API used there. > > Just out of curiosity, which distribution did you choose for > your powerpc server? > > Thanks, Akira > > > > > Thanx, Paul > > > > On Thu, Oct 04, 2018 at 11:01:00PM +0800, Junchang Wang wrote: > >> Hi Paul, > >> > >> Thanks a lot! OSU has approved my application and provisioned a PPC > >> server for me. Now I can start testing code on PPC :-). For the ARM > >> server, the application is still in progress. I'll let you know if > >> there are any updates. > >> > >> > >> Thank! > >> --JunchangOn Tue, Oct 2, 2018 at 1:51 AM Paul E. McKenney > >> <paulmck@xxxxxxxxxxxxx> wrote: > >>> > >>> On Mon, Oct 01, 2018 at 11:23:17AM +0800, Junchang Wang wrote: > >>>> Forward to list perfbook since the last mail was blocked due to the > >>>> non-plain-text issue. > >>>> > >>>> > >>>> Hi Paul and list, > >>>> > >>>> I'm again reading the excellent 15th chapter (Advanced > >>>> Synchronization), which helps me eventually understand how a > >>>> hardware/compiler can optimize--and hence ``break''-- my parallelized > >>>> code :-). > >>>> > >>>> I really want to get my hands dirty and to test my code and samples in > >>>> perfbook on real servers. However, all servers in my lab are x86. I > >>>> wonder if there is any chance that we can have access to architectures > >>>> other than x86 to test code. For example, are there any cloud service > >>>> that provides access to POWER or ARM architecture? > >>> > >>> For POWER access for academic or open-source projects, please see here: > >>> > >>> https://osuosl.org/services/powerdev/request_hosting/ > >>> > >>> There are said to be similar services for ARM, for example: > >>> > >>> https://thenewstack.io/cncf-packet-team-provide-free-infrastructure-cloud-developers > >>> > >>> Please let me know how it goes! > >>> > >>> Thanx, Paul > >>> > >> > > >
