lkml.org 
[lkml]   [2012]   [Nov]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: Problem in Page Cache Replacement
On Wed, Nov 21, 2012 at 05:02:04PM +0800, Fengguang Wu wrote:
> On Wed, Nov 21, 2012 at 04:34:40PM +0800, Jaegeuk Hanse wrote:
> > Cc Fengguang Wu.
> >
> > On 11/21/2012 04:13 PM, metin d wrote:
> > >> Curious. Added linux-mm list to CC to catch more attention. If you run
> > >>echo 1 >/proc/sys/vm/drop_caches does it evict data-1 pages from memory?
> > >I'm guessing it'd evict the entries, but am wondering if we could run any more diagnostics before trying this.
> > >
> > >We regularly use a setup where we have two databases; one gets used frequently and the other one about once a month. It seems like the memory manager keeps unused pages in memory at the expense of frequently used database's performance.
>
> > >My understanding was that under memory pressure from heavily
> > >accessed pages, unused pages would eventually get evicted. Is there
> > >anything else we can try on this host to understand why this is
> > >happening?
>
> We may debug it this way.

Better to add a step

0) run 'page-types -r' to get an initial view of the page cache
status.

Thanks,
Fengguang

> 1) run 'fadvise data-2 0 0 dontneed' to drop data-2 cached pages
> (please double check via /proc/vmstat whether it does the expected work)
>
> 2) run 'page-types -r' with root, to view the page status for the
> remaining pages of data-1
>
> The fadvise tool comes from Andrew Morton's ext3-tools. (source code attached)
> Please compile them with options "-Dlinux -I. -D_GNU_SOURCE -D_FILE_OFFSET_BITS=64 -D_LARGEFILE64_SOURCE"
>
> page-types can be found in the kernel source tree tools/vm/page-types.c
>
> Sorry that sounds a bit twisted.. I do have a patch to directly dump
> page cache status of a user specified file, however it's not
> upstreamed yet.
>
> Thanks,
> Fengguang
>
> > >On Tue 20-11-12 09:42:42, metin d wrote:
> > >>I have two PostgreSQL databases named data-1 and data-2 that sit on the
> > >>same machine. Both databases keep 40 GB of data, and the total memory
> > >>available on the machine is 68GB.
> > >>
> > >>I started data-1 and data-2, and ran several queries to go over all their
> > >>data. Then, I shut down data-1 and kept issuing queries against data-2.
> > >>For some reason, the OS still holds on to large parts of data-1's pages
> > >>in its page cache, and reserves about 35 GB of RAM to data-2's files. As
> > >>a result, my queries on data-2 keep hitting disk.
> > >>
> > >>I'm checking page cache usage with fincore. When I run a table scan query
> > >>against data-2, I see that data-2's pages get evicted and put back into
> > >>the cache in a round-robin manner. Nothing happens to data-1's pages,
> > >>although they haven't been touched for days.
> > >>
> > >>Does anybody know why data-1's pages aren't evicted from the page cache?
> > >>I'm open to all kind of suggestions you think it might relate to problem.
> > > Curious. Added linux-mm list to CC to catch more attention. If you run
> > >echo 1 >/proc/sys/vm/drop_caches
> > > does it evict data-1 pages from memory?
> > >
> > >>This is an EC2 m2.4xlarge instance on Amazon with 68 GB of RAM and no
> > >>swap space. The kernel version is:
> > >>
> > >>$ uname -r
> > >>3.2.28-45.62.amzn1.x86_64
> > >>Edit:
> > >>
> > >>and it seems that I use one NUMA instance, if you think that it can a problem.
> > >>
> > >>$ numactl --hardware
> > >>available: 1 nodes (0)
> > >>node 0 cpus: 0 1 2 3 4 5 6 7
> > >>node 0 size: 70007 MB
> > >>node 0 free: 360 MB
> > >>node distances:
> > >>node 0
> > >> 0: 10

> #include <unistd.h>
> #include <stdlib.h>
> #include <fcntl.h>
> #include <errno.h>
> #include <stdio.h>
> #include <string.h>
>
> #include "fadvise.h"
>
> char *progname;
>
> static void usage(void)
> {
> fprintf(stderr, "Usage: %s filename offset length advice [loops]\n", progname);
> fprintf(stderr, " advice: normal sequential willneed noreuse "
> "dontneed asyncwrite writewait\n");
> exit(1);
> }
>
> int
> main(int argc, char *argv[])
> {
> int c;
> int fd;
> char *sadvice;
> char *filename;
> loff_t offset;
> unsigned long length;
> int advice = 0;
> int ret;
> int loops = 1;
>
> progname = argv[0];
>
> while ((c = getopt(argc, argv, "")) != -1) {
> switch (c) {
> }
> }
>
> if (optind == argc)
> usage();
> filename = argv[optind++];
>
> if (optind == argc)
> usage();
> offset = strtoull(argv[optind++], NULL, 0);
>
> if (optind == argc)
> usage();
> length = strtol(argv[optind++], NULL, 0);
>
> if (optind == argc)
> usage();
> sadvice = argv[optind++];
>
> if (optind != argc)
> loops = strtol(argv[optind++], NULL, 0);
>
> if (optind != argc)
> usage();
>
> if (!strcmp(sadvice, "normal"))
> advice = POSIX_FADV_NORMAL;
> else if (!strcmp(sadvice, "sequential"))
> advice = POSIX_FADV_SEQUENTIAL;
> else if (!strcmp(sadvice, "willneed"))
> advice = POSIX_FADV_WILLNEED;
> else if (!strcmp(sadvice, "noreuse"))
> advice = POSIX_FADV_NOREUSE;
> else if (!strcmp(sadvice, "dontneed"))
> advice = POSIX_FADV_DONTNEED;
> else if (!strcmp(sadvice, "asyncwrite"))
> advice = LINUX_FADV_ASYNC_WRITE;
> else if (!strcmp(sadvice, "writewait"))
> advice = LINUX_FADV_WRITE_WAIT;
> else
> usage();
>
> fd = open(filename, O_RDONLY);
> if (fd < 0) {
> fprintf(stderr, "%s: cannot open `%s': %s\n",
> progname, filename, strerror(errno));
> exit(1);
> }
>
> while (loops--) {
> ret = __posix_fadvise64(fd, offset, length, advice);
> if (ret) {
> fprintf(stderr, "%s: fadvise() failed: %s\n",
> progname, strerror(errno));
> exit(1);
> }
> }
> close(fd);
> exit(0);
> }

> #include <asm/unistd.h>
> #include <sys/errno.h>
>
> #ifndef __NR_fadvise64
> #if defined (__i386__)
> #define __NR_fadvise64 250
> #elif defined(__powerpc__)
> #define __NR_fadvise64 233
> #elif defined(__ia64__)
> #define __NR_fadvise64 1234
> #elif defined(__x86_64__)
> #define __NR_fadvise64 221
> #endif
> #endif
>
> #ifndef LINUX_FADV_ASYNC_WRITE
> #define LINUX_FADV_ASYNC_WRITE 32
> #endif
>
> #ifndef LINUX_FADV_WRITE_WAIT
> #define LINUX_FADV_WRITE_WAIT 33
> #endif
>
> #ifndef __x86_64__
> _syscall5(int,fadvise64, int,fd, long,offset_lo,
> long,offset_hi, size_t,len, int,advice)
> #endif
>
> /* Works by luck on ppc32, fails on ppc64 */
> #if defined(__i386__)
> int __posix_fadvise(int fd, off_t offset, size_t len, int advice)
> {
> return fadvise64(fd, offset, 0, len, advice);
> }
>
> int __posix_fadvise64(int fd, loff_t offset, size_t len, int advice)
> {
> return fadvise64(fd, offset, offset >> 32, len, advice);
> }
> #elif defined(__powerpc64__)
> int __posix_fadvise(int fd, off_t offset, size_t len, int advice)
> {
> return fadvise64(fd, offset, len, advice);
> }
>
> int __posix_fadvise64(int fd, loff_t offset, size_t len, int advice)
> {
> return fadvise64(fd, offset, len, advice);
> }
> #elif defined(__powerpc__)
>
> /*
> * long longs are passed in an odd even register pair on ppc32 so
> * we need to pad before offset
> *
> * Note also the glibc syscall() function for ppc has been broken for
> * 6 argument syscalls until recently (~2.3.1 CVS)
> */
> #define ppc_fadvise64(fd, offset_hi, offset_lo, len, advice) \
> syscall(__NR_fadvise64, fd, 0, offset_hi, offset_lo, len, advice)
>
> int __posix_fadvise(int fd, off_t offset, size_t len, int advice)
> {
> return ppc_fadvise64(fd, 0, offset, len, advice);
> }
>
> /* big endian, akpm. */
> int __posix_fadvise64(int fd, loff_t offset, size_t len, int advice)
> {
> return ppc_fadvise64(fd, (unsigned int)(offset >> 32),
> (unsigned int)(offset & 0xffffffff), len, advice);
> }
> #elif defined(__ia64__)
> int __posix_fadvise(int fd, off_t offset, size_t len, int advice)
> {
> return fadvise64(fd, offset, len, advice);
> }
>
> int __posix_fadvise64(int fd, loff_t offset, size_t len, int advice)
> {
> return fadvise64(fd, offset, len, advice);
> }
> #elif defined(__x86_64__)
> int __posix_fadvise(int fd, off_t offset, size_t len, int advice)
> {
> return -1;
> }
>
> int __posix_fadvise64(int fd, loff_t offset, size_t len, int advice)
> {
> return syscall(__NR_fadvise64, fd, offset, len, advice);
> }
> #endif



\
 
 \ /
  Last update: 2012-11-21 11:01    [W:0.244 / U:0.628 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site