On Thu, Aug 8, 2019 at 12:07 AM René Scharfe <l.s.r@xxxxxx> wrote: > > Am 08.08.19 um 04:35 schrieb Carlo Arenas: > > On Wed, Aug 7, 2019 at 6:03 AM René Scharfe <l.s.r@xxxxxx> wrote: > >> > >> Am 07.08.19 um 11:49 schrieb Carlo Arenas: > >>> was hoping will perform better but it seems that testing can be done > >>> only in windows > >> > >> nedmalloc works on other platforms as well. > > > > I meant[1] it works reliably enough to be useful for performance testing. > > You mentioned being concerned about performance several times and I > wondered why each time. I'd expect no measurable difference between > using a custom global context and the internal one of PCRE2 -- setting > two function pointers surely can't take very long, can it? But > measuring is better than guessing, of course. setting the allocator is not a concern, but using it; it requires an extra indirect function call which is usually not very friendly to caches in our speculative execution CPU world. our implementation also adds the wrapper call overhead, but in this case it is just the "cost of doing business" with PCRE2. compilers had gotten a lot better since (mainly because of C++ and the need for it with virtual methods) but I would rather measure. > > goes without saying that the fact that I am using a virtualbox with 2 > > CPUs running Debian 10 on top of macOS (a macbook pro with 4 cores) > > and the test uses by default 8 threads, doesn't help, > > nedmalloc is supposed to run on macOS as well. the last version has some "fix miscompilations in macOS" fixes that might be relevant, and the version we have in tree says it works in the 32-bit version which latest macOS versions are working hard to deprecate (can't even build for it anymore), eitherway trying to run with a nedmalloc enabled git in macOS is not fun. > > with the only relevant line (for my code) being 7820.19 where it would > > seem it performs almost the same (eventhough just adding NED made it > > initially worst) > > > > note though that the fact there are 20% swings in parts of the code > > that hasn't changed > > or that where explicitly #ifdef out of my code changes doesn't give me > > much confidence, but since the windows guys seem to be using NED by > > default, I am hoping it works better there. > > These measurement results are quite noisy, so I wouldn't trust them too > much. nedmalloc being slower than the one from a recent glibc version > is not very surprising given this statement from its home page, > https://www.nedprod.com/programs/portable/nedmalloc/: > > "Windows 7, Linux 3.x, FreeBSD 8, Mac OS X 10.6 all contain > state-of-the-art allocators and no third party allocator is > likely to significantly improve on them in real world results" > > In particular I don't think that these results justify coupling the use > of nedmalloc to the choice of using a custom global context for PCRE2. neither did I either, the only reason I am holding on fully enabling NED with PCRE2 in my series is just because I wan't to make sure we have identified the bug correctly and we are fixing it (specially since I can't reproduce it, and therefore neither debug it) sorry for not making that clear enough, and as I said before, if we keep seeing segfaults even after v4 then we will have to do that or I might need to do a quick run to the nearest microsoft store hoping for a distracted rep, instead. > I'd expect: > - Without USE_NED_ALLOCATOR: xmalloc() should be used for all > allocations, including for PCRE2. Some special exceptions use > malloc(3) directly, but for most uses we want the consistent > out-of-memory handling that xmalloc() brings. that is already in v4 and would expect to carry it forward. this is also what I had in mind when I said we will need some fixes on top of Dscho version if we give up with these. > - With USE_NED_ALLOCATOR: malloc() and xmalloc() use nedmalloc > behind the scenes and free() is similarly overridden, so all > allocations are affected. > - If USE_NED_ALLOCATOR performs worse than the system allocator on > some system then it's the problem of those that turn on that flag. > > Makes sense? completely, but note also that Dscho version would make the performance impacts of using a custom allocator (if any) affect everyone using PCRE2. Carlo
