Thread: ISM shared memory on solaris
I hope this is the right place to send this.. the FAQ in the distribution mentions http://www.PostgreSQL.org/bugs/bugs.php, which doesn't work. We've found that postgresql wasn't using ISM shared memory on solaris, which theoretically would cost performance. The root cause in our case was that the "solaris" define is not defined by our compilers or by postgresql itself. The patch below simple has it check SHM_SHARE_MMU instead, which should work fine. I verified (with 'pmap') that the database is now using ISM on its shared memory, after this patch was applied. --Josh --- sysv_shmem.c.orig 2002-09-04 13:31:24.000000000 -0700 +++ sysv_shmem.c 2003-10-23 12:52:26.756765000 -0700 @@ -143,7 +143,7 @@ on_shmem_exit(IpcMemoryDelete, Int32GetDatum(shmid)); /* OK, should be able to attach to the segment */ -#if defined(solaris) && defined(__sparc__) +#if defined(SHM_SHARE_MMU) && defined(__sparc__) /* use intimate shared memory on SPARC Solaris */ memAddress = shmat(shmid, 0, SHM_SHARE_MMU); #else @@ -323,8 +323,8 @@ shmid = shmget(NextShmemSegID, sizeof(PGShmemHeader), 0); if (shmid < 0) continue; /* failed: must be some other app's */ - -#if defined(solaris) && defined(__sparc__) + +#if defined(SHM_SHARE_MMU) && defined(__sparc__) /* use intimate shared memory on SPARC Solaris */ memAddress = shmat(shmid, 0, SHM_SHARE_MMU); #else
Yikes! We thought we were already using ISM on Solaris. Would you test the attached patch? It uses _solaris_ rather than SHM_SHARE_MMU in the define test. Does that work too? --------------------------------------------------------------------------- Josh Wilmes wrote: > I hope this is the right place to send this.. the FAQ in the > distribution mentions http://www.PostgreSQL.org/bugs/bugs.php, which > doesn't work. > > We've found that postgresql wasn't using ISM shared memory on solaris, > which theoretically would cost performance. The root cause in our case > was that the "solaris" define is not defined by our compilers or by > postgresql itself. > > The patch below simple has it check SHM_SHARE_MMU instead, which should > work fine. I verified (with 'pmap') that the database is now using ISM > on its shared memory, after this patch was applied. > > --Josh > > > > --- sysv_shmem.c.orig 2002-09-04 13:31:24.000000000 -0700 > +++ sysv_shmem.c 2003-10-23 12:52:26.756765000 -0700 > @@ -143,7 +143,7 @@ > on_shmem_exit(IpcMemoryDelete, Int32GetDatum(shmid)); > > /* OK, should be able to attach to the segment */ > -#if defined(solaris) && defined(__sparc__) > +#if defined(SHM_SHARE_MMU) && defined(__sparc__) > /* use intimate shared memory on SPARC Solaris */ > memAddress = shmat(shmid, 0, SHM_SHARE_MMU); > #else > @@ -323,8 +323,8 @@ > shmid = shmget(NextShmemSegID, sizeof(PGShmemHeader), 0); > if (shmid < 0) > continue; /* failed: must > be some other app's */ > - > -#if defined(solaris) && defined(__sparc__) > + > +#if defined(SHM_SHARE_MMU) && defined(__sparc__) > /* use intimate shared memory on SPARC Solaris */ > memAddress = shmat(shmid, 0, SHM_SHARE_MMU); > #else > > > > ---------------------------(end of broadcast)--------------------------- > TIP 3: if posting/reading through Usenet, please send an appropriate > subscribe-nomail command to majordomo@postgresql.org so that your > message can get through to the mailing list cleanly > -- Bruce Momjian | http://candle.pha.pa.us pgman@candle.pha.pa.us | (610) 359-1001 + If your life is a hard drive, | 13 Roberts Road + Christ can be your backup. | Newtown Square, Pennsylvania 19073 Index: src/backend/port/sysv_shmem.c =================================================================== RCS file: /cvsroot/pgsql-server/src/backend/port/sysv_shmem.c,v retrieving revision 1.21 diff -c -c -r1.21 sysv_shmem.c *** src/backend/port/sysv_shmem.c 13 Oct 2003 22:47:15 -0000 1.21 --- src/backend/port/sysv_shmem.c 24 Oct 2003 15:46:03 -0000 *************** *** 133,139 **** on_shmem_exit(IpcMemoryDelete, Int32GetDatum(shmid)); /* OK, should be able to attach to the segment */ ! #if defined(solaris) && defined(__sparc__) /* use intimate shared memory on SPARC Solaris */ memAddress = shmat(shmid, 0, SHM_SHARE_MMU); #else --- 133,139 ---- on_shmem_exit(IpcMemoryDelete, Int32GetDatum(shmid)); /* OK, should be able to attach to the segment */ ! #if defined(__solaris__) && defined(__sparc__) /* use intimate shared memory on SPARC Solaris */ memAddress = shmat(shmid, 0, SHM_SHARE_MMU); #else *************** *** 352,358 **** hdr = (PGShmemHeader *) shmat(*shmid, UsedShmemSegAddr, ! #if defined(solaris) && defined(__sparc__) /* use intimate shared memory on Solaris */ SHM_SHARE_MMU #else --- 352,358 ---- hdr = (PGShmemHeader *) shmat(*shmid, UsedShmemSegAddr, ! #if defined(__solaris__) && defined(__sparc__) /* use intimate shared memory on Solaris */ SHM_SHARE_MMU #else
Josh Wilmes wrote: > Nope, __solaris__ is not defined on our system either. I thought our configure defined __portname__ for every platform, but I don't see that anywhere, so it seems we rely on the compiler to supply defines for the cpu and OS. Does src/tools/ccsym show you your defines? I would like to have something that identifies Solaris rather than something that checks for ISM so that if the ISM define isn't found, we throw an error and we hear about it. -- Bruce Momjian | http://candle.pha.pa.us pgman@candle.pha.pa.us | (610) 359-1001 + If your life is a hard drive, | 13 Roberts Road + Christ can be your backup. | Newtown Square, Pennsylvania 19073
Bruce Momjian wrote: > Josh Wilmes wrote: > >>Nope, __solaris__ is not defined on our system either. > > > I thought our configure defined __portname__ for every platform, but I > don't see that anywhere, so it seems we rely on the compiler to supply > defines for the cpu and OS. > > Does src/tools/ccsym show you your defines? I would like to have > something that identifies Solaris rather than something that checks for > ISM so that if the ISM define isn't found, we throw an error and we hear > about it. That would be preferable- i didn't know what was safe to assume would always be defined. ccsym is pretty neat. Here's what it shows (gcc) __GNUC__=2 __GNUC_MINOR__=95 sparc sun unix __svr4__ __SVR4 __sparc__ __sun__ __unix__ __svr4__ __SVR4 __sparc __sun __unix system=unix system=svr4 __GCC_NEW_VARARGS__ cpu=sparc machine=sparc
Nope, __solaris__ is not defined on our system either. --Josh Bruce Momjian wrote: > Yikes! We thought we were already using ISM on Solaris. > > Would you test the attached patch? It uses _solaris_ rather than > SHM_SHARE_MMU in the define test. Does that work too? > > --------------------------------------------------------------------------- > > Josh Wilmes wrote: > >>I hope this is the right place to send this.. the FAQ in the >>distribution mentions http://www.PostgreSQL.org/bugs/bugs.php, which >>doesn't work. >> >>We've found that postgresql wasn't using ISM shared memory on solaris, >>which theoretically would cost performance. The root cause in our case >>was that the "solaris" define is not defined by our compilers or by >>postgresql itself. >> >>The patch below simple has it check SHM_SHARE_MMU instead, which should >>work fine. I verified (with 'pmap') that the database is now using ISM >> on its shared memory, after this patch was applied. >> >>--Josh >> >> >> >>--- sysv_shmem.c.orig 2002-09-04 13:31:24.000000000 -0700 >>+++ sysv_shmem.c 2003-10-23 12:52:26.756765000 -0700 >>@@ -143,7 +143,7 @@ >> on_shmem_exit(IpcMemoryDelete, Int32GetDatum(shmid)); >> >> /* OK, should be able to attach to the segment */ >>-#if defined(solaris) && defined(__sparc__) >>+#if defined(SHM_SHARE_MMU) && defined(__sparc__) >> /* use intimate shared memory on SPARC Solaris */ >> memAddress = shmat(shmid, 0, SHM_SHARE_MMU); >> #else >>@@ -323,8 +323,8 @@ >> shmid = shmget(NextShmemSegID, sizeof(PGShmemHeader), 0); >> if (shmid < 0) >> continue; /* failed: must >>be some other app's */ >>- >>-#if defined(solaris) && defined(__sparc__) >>+ >>+#if defined(SHM_SHARE_MMU) && defined(__sparc__) >> /* use intimate shared memory on SPARC Solaris */ >> memAddress = shmat(shmid, 0, SHM_SHARE_MMU); >> #else >> >> >> >>---------------------------(end of broadcast)--------------------------- >>TIP 3: if posting/reading through Usenet, please send an appropriate >> subscribe-nomail command to majordomo@postgresql.org so that your >> message can get through to the mailing list cleanly >> > > > > ------------------------------------------------------------------------ > > Index: src/backend/port/sysv_shmem.c > =================================================================== > RCS file: /cvsroot/pgsql-server/src/backend/port/sysv_shmem.c,v > retrieving revision 1.21 > diff -c -c -r1.21 sysv_shmem.c > *** src/backend/port/sysv_shmem.c 13 Oct 2003 22:47:15 -0000 1.21 > --- src/backend/port/sysv_shmem.c 24 Oct 2003 15:46:03 -0000 > *************** > *** 133,139 **** > on_shmem_exit(IpcMemoryDelete, Int32GetDatum(shmid)); > > /* OK, should be able to attach to the segment */ > ! #if defined(solaris) && defined(__sparc__) > /* use intimate shared memory on SPARC Solaris */ > memAddress = shmat(shmid, 0, SHM_SHARE_MMU); > #else > --- 133,139 ---- > on_shmem_exit(IpcMemoryDelete, Int32GetDatum(shmid)); > > /* OK, should be able to attach to the segment */ > ! #if defined(__solaris__) && defined(__sparc__) > /* use intimate shared memory on SPARC Solaris */ > memAddress = shmat(shmid, 0, SHM_SHARE_MMU); > #else > *************** > *** 352,358 **** > > hdr = (PGShmemHeader *) shmat(*shmid, > UsedShmemSegAddr, > ! #if defined(solaris) && defined(__sparc__) > /* use intimate shared memory on Solaris */ > SHM_SHARE_MMU > #else > --- 352,358 ---- > > hdr = (PGShmemHeader *) shmat(*shmid, > UsedShmemSegAddr, > ! #if defined(__solaris__) && defined(__sparc__) > /* use intimate shared memory on Solaris */ > SHM_SHARE_MMU > #else
OK, attached patch applied. I tested for 'sun' rather than 'solaris'. --------------------------------------------------------------------------- Josh Wilmes wrote: > > Bruce Momjian wrote: > > > Josh Wilmes wrote: > > > >>Nope, __solaris__ is not defined on our system either. > > > > > > I thought our configure defined __portname__ for every platform, but I > > don't see that anywhere, so it seems we rely on the compiler to supply > > defines for the cpu and OS. > > > > Does src/tools/ccsym show you your defines? I would like to have > > something that identifies Solaris rather than something that checks for > > ISM so that if the ISM define isn't found, we throw an error and we hear > > about it. > > That would be preferable- i didn't know what was safe to assume would > always be defined. > > ccsym is pretty neat. Here's what it shows (gcc) > > __GNUC__=2 > __GNUC_MINOR__=95 > sparc > sun > unix > __svr4__ > __SVR4 > __sparc__ > __sun__ > __unix__ > __svr4__ > __SVR4 > __sparc > __sun > __unix > system=unix > system=svr4 > __GCC_NEW_VARARGS__ > cpu=sparc > machine=sparc > > > > ---------------------------(end of broadcast)--------------------------- > TIP 7: don't forget to increase your free space map settings > -- Bruce Momjian | http://candle.pha.pa.us pgman@candle.pha.pa.us | (610) 359-1001 + If your life is a hard drive, | 13 Roberts Road + Christ can be your backup. | Newtown Square, Pennsylvania 19073 Index: src/backend/port/sysv_shmem.c =================================================================== RCS file: /cvsroot/pgsql-server/src/backend/port/sysv_shmem.c,v retrieving revision 1.21 diff -c -c -r1.21 sysv_shmem.c *** src/backend/port/sysv_shmem.c 13 Oct 2003 22:47:15 -0000 1.21 --- src/backend/port/sysv_shmem.c 26 Oct 2003 04:52:33 -0000 *************** *** 133,139 **** on_shmem_exit(IpcMemoryDelete, Int32GetDatum(shmid)); /* OK, should be able to attach to the segment */ ! #if defined(solaris) && defined(__sparc__) /* use intimate shared memory on SPARC Solaris */ memAddress = shmat(shmid, 0, SHM_SHARE_MMU); #else --- 133,139 ---- on_shmem_exit(IpcMemoryDelete, Int32GetDatum(shmid)); /* OK, should be able to attach to the segment */ ! #if defined(sun) && defined(__sparc__) /* use intimate shared memory on SPARC Solaris */ memAddress = shmat(shmid, 0, SHM_SHARE_MMU); #else *************** *** 352,358 **** hdr = (PGShmemHeader *) shmat(*shmid, UsedShmemSegAddr, ! #if defined(solaris) && defined(__sparc__) /* use intimate shared memory on Solaris */ SHM_SHARE_MMU #else --- 352,358 ---- hdr = (PGShmemHeader *) shmat(*shmid, UsedShmemSegAddr, ! #if defined(sun) && defined(__sparc__) /* use intimate shared memory on Solaris */ SHM_SHARE_MMU #else
Please use new patch. This uses __sun__ rather than sun. --------------------------------------------------------------------------- Josh Wilmes wrote: > > Bruce Momjian wrote: > > > Josh Wilmes wrote: > > > >>Nope, __solaris__ is not defined on our system either. > > > > > > I thought our configure defined __portname__ for every platform, but I > > don't see that anywhere, so it seems we rely on the compiler to supply > > defines for the cpu and OS. > > > > Does src/tools/ccsym show you your defines? I would like to have > > something that identifies Solaris rather than something that checks for > > ISM so that if the ISM define isn't found, we throw an error and we hear > > about it. > > That would be preferable- i didn't know what was safe to assume would > always be defined. > > ccsym is pretty neat. Here's what it shows (gcc) > > __GNUC__=2 > __GNUC_MINOR__=95 > sparc > sun > unix > __svr4__ > __SVR4 > __sparc__ > __sun__ > __unix__ > __svr4__ > __SVR4 > __sparc > __sun > __unix > system=unix > system=svr4 > __GCC_NEW_VARARGS__ > cpu=sparc > machine=sparc > > > > ---------------------------(end of broadcast)--------------------------- > TIP 7: don't forget to increase your free space map settings > -- Bruce Momjian | http://candle.pha.pa.us pgman@candle.pha.pa.us | (610) 359-1001 + If your life is a hard drive, | 13 Roberts Road + Christ can be your backup. | Newtown Square, Pennsylvania 19073 Index: src/backend/port/sysv_shmem.c =================================================================== RCS file: /cvsroot/pgsql-server/src/backend/port/sysv_shmem.c,v retrieving revision 1.21 diff -c -c -r1.21 sysv_shmem.c *** src/backend/port/sysv_shmem.c 13 Oct 2003 22:47:15 -0000 1.21 --- src/backend/port/sysv_shmem.c 26 Oct 2003 04:52:33 -0000 *************** *** 133,139 **** on_shmem_exit(IpcMemoryDelete, Int32GetDatum(shmid)); /* OK, should be able to attach to the segment */ ! #if defined(solaris) && defined(__sparc__) /* use intimate shared memory on SPARC Solaris */ memAddress = shmat(shmid, 0, SHM_SHARE_MMU); #else --- 133,139 ---- on_shmem_exit(IpcMemoryDelete, Int32GetDatum(shmid)); /* OK, should be able to attach to the segment */ ! #if defined(__sun__) && defined(__sparc__) /* use intimate shared memory on SPARC Solaris */ memAddress = shmat(shmid, 0, SHM_SHARE_MMU); #else *************** *** 352,358 **** hdr = (PGShmemHeader *) shmat(*shmid, UsedShmemSegAddr, ! #if defined(solaris) && defined(__sparc__) /* use intimate shared memory on Solaris */ SHM_SHARE_MMU #else --- 352,358 ---- hdr = (PGShmemHeader *) shmat(*shmid, UsedShmemSegAddr, ! #if defined(__sun__) && defined(__sparc__) /* use intimate shared memory on Solaris */ SHM_SHARE_MMU #else
Bruce Momjian <pgman@candle.pha.pa.us> writes:
> ! #if defined(sun) && defined(__sparc__)
> /* use intimate shared memory on SPARC Solaris */
> memAddress = shmat(shmid, 0, SHM_SHARE_MMU);
> #else
I think this is going in the wrong direction. Why isn't the code just
#if defined(SHM_SHARE_MMU)
/* use intimate shared memory on Solaris */
memAddress = shmat(shmid, 0, SHM_SHARE_MMU);
#else
If the symbol is available I think we probably want to use it. It is an
O/S issue, not a hardware issue, and so the test on __sparc__ seems
quite wrongheaded ...
regards, tom lane
Tom Lane wrote: > Bruce Momjian <pgman@candle.pha.pa.us> writes: > > > ! #if defined(sun) && defined(__sparc__) > > /* use intimate shared memory on SPARC Solaris */ > > memAddress = shmat(shmid, 0, SHM_SHARE_MMU); > > #else > > I think this is going in the wrong direction. Why isn't the code just > > #if defined(SHM_SHARE_MMU) > /* use intimate shared memory on Solaris */ > memAddress = shmat(shmid, 0, SHM_SHARE_MMU); > #else > > If the symbol is available I think we probably want to use it. It is an > O/S issue, not a hardware issue, and so the test on __sparc__ seems > quite wrongheaded ... What I was hoping to do with the define test was to throw an error if we don't find intimate shared memory on Solaris, but the define doesn't work fir i386/Solaris so we are probably better going with the define as you suggest --- I just hope we don't fail to include a file and somehow miss it on some version of Solaris. Change applied. -- Bruce Momjian | http://candle.pha.pa.us pgman@candle.pha.pa.us | (610) 359-1001 + If your life is a hard drive, | 13 Roberts Road + Christ can be your backup. | Newtown Square, Pennsylvania 19073
Seems like the BEST case would be to have a configure test verify that it works and define something if it does, but i don't know what such a test would look like. --Josh Bruce Momjian wrote: > Tom Lane wrote: > >>Bruce Momjian <pgman@candle.pha.pa.us> writes: >> >> >>>! #if defined(sun) && defined(__sparc__) >>> /* use intimate shared memory on SPARC Solaris */ >>> memAddress = shmat(shmid, 0, SHM_SHARE_MMU); >>> #else >> >>I think this is going in the wrong direction. Why isn't the code just >> >>#if defined(SHM_SHARE_MMU) >> /* use intimate shared memory on Solaris */ >> memAddress = shmat(shmid, 0, SHM_SHARE_MMU); >>#else >> >>If the symbol is available I think we probably want to use it. It is an >>O/S issue, not a hardware issue, and so the test on __sparc__ seems >>quite wrongheaded ... > > > What I was hoping to do with the define test was to throw an error if we > don't find intimate shared memory on Solaris, but the define doesn't > work fir i386/Solaris so we are probably better going with the define as > you suggest --- I just hope we don't fail to include a file and somehow > miss it on some version of Solaris. > > Change applied. >