On Jan 4, 8:35 am, Cactus <rieman...@googlemail.com> wrote: > On Jan 4, 3:26 am, Case Vanhorsen <cas...@gmail.com> wrote: > > > > > On Sun, Jan 3, 2010 at 12:33 PM, Cactus <rieman...@googlemail.com> wrote: > > > > On Jan 3, 8:15 pm, Case Vanhorsen <cas...@gmail.com> wrote: > > >> Hello, > > > >> I discovered an interesting memory allocation behavior on Windows vs. > > >> Linux. I was testing GMPY on 64-bit Windows when I stumbled into this. > > >> GMPY replaces the native MPIR memory allocation routines with Python's > > >> memory allocator. If I enable debugging in GMPY, I get a trace of all > > >> the memory allocation calls. When I ran the following: > > > >> python -mtimeit -n 1 -r 1 -s "import > > >> gmpy;gmpy.mpz(3)**(2**27);gmpy.set_debug(1)" "a=a*a" 2>temp.txt > > > >> and look at the output saved in temp.txt, I see that Linux generated > > >> approximately 34 memory manager calls but Windows generates over > > >> 100,000 calls. Most of the Windows allocations are for small (<8K) > > >> chunks of memory while all the Linux requests are for more than 64K. > > >> The performance between Linux and Windows is similar. Could it be that > > >> Windows is not using alloca? > > > >> I also think I found another memory allocation bug. If I run the above > > >> multiplication repeatedly: > > > >> python -mtimeit -s "import > > >> gmpy;gmpy.mpz(3)**(2**27);gmpy.set_debug(1)" "a=a*a" 2>temp.txt > > > >> it will eventually crash. In looking at the debug output, I see a > > >> request to allocate 18446744073709498400 bytes of memory. > > > >> I tested with both MPIR 1.2.2 and 1.3.0 and get similar behavior. I'm > > >> using a custom version of GMPY with some fixes for size_t vs. long > > >> issues that hasn't been committed but I will try to commit those > > >> changes later today. > > > > Hi Case > > > > That is _very_ useful information and may explain why Windows > > > performance lgas that on Linux. > > > > A long time ago - in GMP days - I had to turn of the use of alloca as > > > I kept getting crashes if I used it. > > > > I have not tried switching it on in MPIR but I will certainly look at > > > this again. > > > > But _alloca is now deprecated on Windows and its replacement requires > > > a 'free' procedure that alloca doesn't need. It may hence be quite > > > difficult to take advantage of this in future. It all depends on how > > > GMP/MPIR use alloca. > > > > Thanks for the debugging! > > > > Brian > > > Some additional information. > > > MPIR 1.2.2 generates approximately 172,000 memory allocator calls. > > MPIR 1.3.0 generates approximately 221,000 memory allocator calls. > > > The count includes both mp_allocate and mp_free. > > > Performance numbers: > > Windows x64, MPIR 1.2.2: 3.35 seconds > > Windows x64, MPIR 1.3.0, 3.76 seconds > > Linux, MPIR 1.3.0, 3.05 seconds > > > Processor is a Core2. > > > Regarding the mp_allocate failure: it looks like it occurs when trying > > to allocate space for a number larger that 2^32 bits long. > > Does this allocation failure only occur on Windows? > > The MPIR settings for memory management on Windows are: > > #define HAVE_ALLOCA 1 > #undef HAVE_ALLOCA_H > #undef WANT_TMP_ALLOCA > #undef WANT_TMP_DEBUG > #undef WANT_TMP_NOTREENTRANT > #define WANT_TMP_REENTRANT 1 > > What are the normal settings used on a Linux/GCC build?
The memory allocation calls in MPIR are of trhe form void *memory(pointer, size_t) Since size_t is a 32-bit int on Windows, this will fail if an attempt is made to obtain 2^32 or more _bytes_. But it should be ok for 2^32 bits. Brian -- You received this message because you are subscribed to the Google Groups "mpir-devel" group. To post to this group, send email to mpir-de...@googlegroups.com. To unsubscribe from this group, send email to mpir-devel+unsubscr...@googlegroups.com. For more options, visit this group at http://groups.google.com/group/mpir-devel?hl=en.