Transparent Hugepage Support #29

http://git.kernel.org/?p=linux/kernel/git/andrea/aa.git;a=shortlog

first: git clone git://git.kernel.org/pub/scm/linux/kernel/git/andrea/aa.git
or first: git clone --reference linux-2.6 git://git.kernel.org/pub/scm/linux/kernel/git/andrea/aa.git
later: git fetch; git checkout -f origin/master

The tree is rebased and git pull won't work.

http://www.kernel.org/pub/linux/kernel/people/andrea/patches/v2.6/2.6.35/transparent_hugepage-29/
http://www.kernel.org/pub/linux/kernel/people/andrea/patches/v2.6/2.6.35/transparent_hugepage-29.gz

Diff #28 -> #29:

 THP-disable-on-small-systems    |   41 ++++++++

Disable THP on <500M systems, from Rik.

 compaction-kswapd               |  194 ++++++++++++++++++++++++++++++++++++++++

Add a compaction mode for kswapd to fully replace obsolete blind lumpy reclaim.

 free_pages-count                |   60 ++++++++++++
 free_pages-drain_all_pages      |   62 ++++++++++++
 free_pages-vmstat               |  156 ++++++++++++++++++++++++++++++++

Make free page statistics more accurate from Mel.

 ksmd-khugepaged-freeze          |   97 ++++++++++++++++++++

kswapd uses set_freezing, so shall ksmd and khugepaged.

 transparent-hugepage-nr_rotated |   36 +++++++
 transparent-hugepage-stat       |  166 ++++++++++++++++++++++++++++++++++

Fix inactive/active stats and nr_rotated, from Rik.

 Documentation/vm/transhuge.txt        |  283 ++++
 arch/alpha/include/asm/mman.h         |    2 
 arch/mips/include/asm/mman.h          |    2 
 arch/parisc/include/asm/mman.h        |    2 
 arch/powerpc/mm/gup.c                 |   12 
 arch/x86/include/asm/kvm_host.h       |    1 
 arch/x86/include/asm/paravirt.h       |   23 
 arch/x86/include/asm/paravirt_types.h |    6 
 arch/x86/include/asm/pgtable-2level.h |    9 
 arch/x86/include/asm/pgtable-3level.h |   23 
 arch/x86/include/asm/pgtable.h        |  149 ++
 arch/x86/include/asm/pgtable_64.h     |   28 
 arch/x86/include/asm/pgtable_types.h  |    3 
 arch/x86/kernel/paravirt.c            |    3 
 arch/x86/kernel/vm86_32.c             |    1 
 arch/x86/kvm/mmu.c                    |   68 -
 arch/x86/kvm/paging_tmpl.h            |    6 
 arch/x86/mm/gup.c                     |   28 
 arch/x86/mm/pgtable.c                 |   66 +
 arch/xtensa/include/asm/mman.h        |    2 
 fs/Kconfig                            |    2 
 fs/exec.c                             |   44 
 fs/proc/meminfo.c                     |   14 
 fs/proc/page.c                        |   14 
 include/asm-generic/mman-common.h     |    2 
 include/asm-generic/pgtable.h         |  130 +
 include/linux/compaction.h            |   13 
 include/linux/gfp.h                   |   14 
 include/linux/huge_mm.h               |  151 ++
 include/linux/khugepaged.h            |   66 +
 include/linux/ksm.h                   |   20 
 include/linux/kvm_host.h              |    4 
 include/linux/memory_hotplug.h        |   14 
 include/linux/mm.h                    |  114 +
 include/linux/mm_inline.h             |   19 
 include/linux/mm_types.h              |    3 
 include/linux/mmu_notifier.h          |   66 +
 include/linux/mmzone.h                |    1 
 include/linux/page-flags.h            |   36 
 include/linux/res_counter.h           |   12 
 include/linux/rmap.h                  |   33 
 include/linux/sched.h                 |    1 
 include/linux/swap.h                  |    2 
 include/linux/vmstat.h                |    4 
 kernel/fork.c                         |   12 
 kernel/futex.c                        |   55 
 mm/Kconfig                            |   38 
 mm/Makefile                           |    1 
 mm/compaction.c                       |   48 
 mm/huge_memory.c                      | 2212 ++++++++++++++++++++++++++++++++++
 mm/hugetlb.c                          |   69 -
 mm/ksm.c                              |   86 -
 mm/madvise.c                          |    8 
 mm/memcontrol.c                       |  169 +-
 mm/memory-failure.c                   |    2 
 mm/memory.c                           |  241 +++
 mm/memory_hotplug.c                   |   14 
 mm/mempolicy.c                        |   14 
 mm/migrate.c                          |   22 
 mm/mincore.c                          |    7 
 mm/mmap.c                             |   57 
 mm/mmu_notifier.c                     |   20 
 mm/mprotect.c                         |   20 
 mm/mremap.c                           |    8 
 mm/oom_kill.c                         |    1 
 mm/page_alloc.c                       |   58 
 mm/pagewalk.c                         |    1 
 mm/rmap.c                             |  228 ++-
 mm/sparse.c                           |    4 
 mm/swap.c                             |  117 +
 mm/swap_state.c                       |    6 
 mm/swapfile.c                         |    2 
 mm/vmscan.c                           |  102 -
 mm/vmstat.c                           |   31 
 virt/kvm/iommu.c                      |    2 
 virt/kvm/kvm_main.c                   |   56 
 76 files changed, 4680 insertions(+), 527 deletions(-)

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Transparent Hugepage Support worst case macro benchmark: kernel build.

host: 24-way SMP (12cores, 2 sockets) 16G RAM
guest: 24-way SMP (24 vcpus) 15G RAM

- Same kernel in guest and host: aa.git tag THP-29 (2.6.35 based)
http://git.kernel.org/?p=linux/kernel/git/andrea/aa.git;a=summary
(see linux-mm for more info)
- Same gcc (and userland) patched with the patch at the end.
- Same kernel source, same .config, on tmpfs (tmpfs only eliminates the measurement error across different runs but ext4 leads to the same average results).
- khugepaged default settings (khugepaged taking 0% CPU)
- no glibc align tweak (that would improve performance a little further with THP always)

The measurement also includes the "make clean", and the full "make -j32" that includes lots of other time consuming operations not getting any benefit from transparent hugepages. If this was pure "gcc" the percentage speedup would be much higher than this. This is a very real life workload that we run on a daily basis, not a microbenchmark at all.

Kernel build on bare metal (note the dTLB-load-misses):

====== build ======
#!/bin/bash
make clean >/dev/null; make -j32 >/dev/null
===================
perf stat -e cycles -e instructions -e dtlb-loads -e dtlb-load-misses --repeat 3 ./build
===================

THP always host (fastest base result)

Performance counter stats for './build' (3 runs):

      4420734012848  cycles                     ( +-   0.007% )
      2692414418384  instructions             #      0.609 IPC     ( +-   0.000% )
       696638665612  dTLB-loads                 ( +-   0.001% )
         2982343758  dTLB-load-misses           ( +-   0.051% )

       83.855147696  seconds time elapsed   ( +-   0.058% )

THP never host (slowdown 4.06%)

Performance counter stats for './build' (3 runs):

      4599325985460  cycles                     ( +-   0.013% )
      2747874065083  instructions             #      0.597 IPC     ( +-   0.000% )
       710631792376  dTLB-loads                 ( +-   0.000% )
         4425816093  dTLB-load-misses           ( +-   0.039% )

       87.260443531  seconds time elapsed   ( +-   0.075% )

Kernel build on KVM powered guest:

=======
time (make clean; make -j32) >/dev/null
=======

THP always guest, EPT on + THP always host (slowdown 5.67%)

NOTE: the total KVM virtualization slowdown for the kernel build with THP always in guest and host compared with bare metal with THP never (like current upstream) is only 1.54%.

real 1m28.612s -> 88.612 seconds
user 26m13.862s
sys 2m11.376s

THP never guest, EPT on + THP always host (slowdown 12.71%)

real 1m34.516s -> 94.516 seconds
user 26m52.929s
sys 3m35.509s

THP never guest, EPT on + THP never host (slowdown 24.81%)

real 1m44.663s -> 104.663 seconds
user 28m13.382s
sys 5m39.373s

THP always guest, EPT off + THP always host (slowdown 198.33%)

real 4m10.166s -> 250.166 seconds
user 41m5.674s
sys 47m37.671s

THP never guest, EPT off + THP always host (slowdown 254.43%)

real 4m57.211s -> 297.211 seconds
user 53m44.302s
sys 53m21.600s

THP never guest, EPT off + THP never host (slowdown 260.15%)

real 5m2.006s -> 302.006 seconds
user 53m25.876s
sys 53m32.649s

This is trivial to reproduce, you can try yourself with aa.git and the below gcc patch. The results are similar to what I got with NPT some time ago with a less complete benchmark (no gcc patched in guest) and only 4 cores.

After this worst case macro benchmark, I'll go ahead with more optimal benchmarks (with bigger memory footprint and longer-living tasks not quitting so fast and not including make -j32/make clean etc...) and I expect more pronounced speedups, like the qemu-kvm translate.o gcc build (the file with the automatic generated .c source for JIT emulation), Java etc... And I'll include these results and all other results I'll be getting, in my KVM Forum 2010 talk on Transparent Hugepage Support next week in Boston.

Thanks!
Andrea

@@ -450,6 +450,11 @@
 #define BITMAP_SIZE(Num_objects) \
   (CEIL ((Num_objects), HOST_BITS_PER_LONG) * sizeof(long))
 
+#ifdef __x86_64__
+#define HPAGE_SIZE (2*1024*1024)
+#define GGC_QUIRE_SIZE 512
+#endif
+
 /* Allocate pages in chunks of this size, to throttle calls to memory
    allocation routines.  The first page is used, the rest go onto the
    free list.  This cannot be larger than HOST_BITS_PER_INT for the

@@ -457,7 +462,7 @@
    can override this by defining GGC_QUIRE_SIZE explicitly.  */
 #ifndef GGC_QUIRE_SIZE
 # ifdef USING_MMAP
-#  define GGC_QUIRE_SIZE 256
+#  define GGC_QUIRE_SIZE 512
 # else
 #  define GGC_QUIRE_SIZE 16
 # endif

@@ -654,6 +659,23 @@
 #ifdef HAVE_MMAP_ANON
   char *page = (char *) mmap (pref, size, PROT_READ | PROT_WRITE,
 			      MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+#ifdef HPAGE_SIZE
+  if (!(size & (HPAGE_SIZE-1)) &&
+      page != (char *) MAP_FAILED && (size_t) page & (HPAGE_SIZE-1)) {
+	  char *old_page;
+	  munmap(page, size);
+	  page = (char *) mmap (pref, size + HPAGE_SIZE-1,
+				PROT_READ | PROT_WRITE,
+				MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+	  old_page = page;
+	  page = (char *) (((size_t)page + HPAGE_SIZE-1)
+			   & ~(HPAGE_SIZE-1));
+	  if (old_page != page)
+		  munmap(old_page, page-old_page);
+	  if (page != old_page + HPAGE_SIZE-1)
+		  munmap(page+size, old_page+HPAGE_SIZE-1-page);
+  }
+#endif
 #endif
 #ifdef HAVE_MMAP_DEV_ZERO
   char *page = (char *) mmap (pref, size, PROT_READ | PROT_WRITE,

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Andrea,

I have been tracking & testing this for a while. It looks very much like Linus wants to merge (or a big chunk of it):

git://git.kernel.org/pub/scm/linux/kernel/git/npiggin/linux-npiggin.git vfs-scale-working

When I pull an aa-29 on top of this there is one conflict in vmscan.c that git cannot resolve. It would be really nice to have a version of the THS patchs that can be tested with the vfs scaling tree.

TIA
Ed Tomlinson

Hi Ed,

please try:

zcat transparent-hugepage-29.gz |patch -p1

It didn't spawn rejects here, maybe something's wrong in some intra-patch or git is too picky. If it doesn't build let me know and I will look into it, thanks!

Andrea

On a tree with:

commit 9d9cec888c3c4a2d5b5f01c3f3984fc4bc602681
Merge: dd6c927 9edd35f
Author: Ed <[email protected]>
Date: Wed Aug 4 15:18:53 2010 -0400

Merge branch 'vfs-scale-working' of git://git.kernel.org/pub/scm/linux/kernel/git/npiggin/linux-npiggin into vfs-scaling

merged on top of

commit 9fe6206f400646a2322096b56c59891d530e8d51

Author: Linus Torvalds <[email protected]>
Date:   Sun Aug 1 15:11:14 2010 -0700

    Linux 2.6.35

doing: bzcat ../transparent-hugepage-29.bz2 | patch -p1

I get one reject:

patching file mm/vmscan.c
Hunk #3 succeeded at 612 (offset 43 lines).
Hunk #4 succeeded at 968 (offset 43 lines).
Hunk #5 succeeded at 1026 (offset 43 lines).
Hunk #6 succeeded at 1063 (offset 43 lines).
Hunk #7 succeeded at 1168 (offset 43 lines).
Hunk #8 succeeded at 1179 (offset 43 lines).
Hunk #9 succeeded at 1198 (offset 43 lines).
Hunk #10 succeeded at 1236 (offset 43 lines).
Hunk #11 succeeded at 1263 (offset 43 lines).
Hunk #12 succeeded at 1333 (offset 43 lines).
Hunk #13 succeeded at 1401 (offset 43 lines).
Hunk #14 succeeded at 1658 (offset 43 lines).
Hunk #15 FAILED at 1629.
Hunk #16 succeeded at 2112 with fuzz 2 (offset 51 lines). Hunk #17 FAILED at 2093.
Hunk #18 succeeded at 2169 (offset 46 lines).
2 out of 18 hunks FAILED -- saving rejects to file mm/vmscan.c.rej patching file mm/vmstat.c

---
cat mm/vmscan.c.rej

@@ -1629,8 +1586,6 @@
 
        get_scan_count(zone, sc, nr, priority);
 
-       set_lumpy_reclaim_mode(priority, sc);
-
        while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
                                        nr[LRU_INACTIVE_FILE]) {
                for_each_evictable_lru(l) {

@@ -2093,9 +2049,25 @@
                                                lru_pages);
                        sc.nr_reclaimed += reclaim_state->reclaimed_slab;
                        total_scanned += sc.nr_scanned;
+                       compaction = 0;
+
+                       if (order &&
+                           zone_watermark_ok(zone, 0,
+                                              high_wmark_pages(zone),
+                                             end_zone, 0) &&
+                           !zone_watermark_ok(zone, order,
+                                              high_wmark_pages(zone),
+                                              end_zone, 0)) {
+                               compact_zone_order(zone,
+                                                  order,
+                                                  sc.gfp_mask,
+                                                  COMPACT_MODE_KSWAPD);
+                               compaction = 1;
+                       }
+
                        if (zone->all_unreclaimable)
                                continue;
-                       if (nr_slab == 0 &&
+                       if (!compaction && nr_slab == 0 &&
                            zone->pages_scanned >= (zone_reclaimable_pages(zone) * 6))
                                zone->all_unreclaimable = 1;
                        /*

--

Ideas?
Ed Tomlinson

Well I didn't realize you also merged 2.6.35 on top of it. This is no standard tree, so I suggest you to checkout -f origin/master of Nick's tree until he updates to 2.6.35 himself (then I can port to it).

Nick,

Any chance of the above happening? I'd love to test both the vfs scaling changes and Andrea's much improved hugepage support together on top of .35

TIA
Ed Tomlinson

I was going to rebase to 2.6.35 sometime, but it may have to wait until I get back from LSF/MM if I don't get time this week.

Try vfs-scale branch now. Cleaned up and rebased on 2.6.35 although it is not as well tested as I would like but I won't have any more time on it until next week.

vfs-scale-working retained history of fixes. I'll start another vfs-scale-working-2.6.36 branch for new fixes.

Thanks,
Nick

Add hugepage statistics to per-node sysfs meminfo

---
 drivers/base/node.c |   21 ++++++++++++++++++---
 1 files changed, 18 insertions(+), 3 deletions(-)

@@ -98,7 +98,11 @@ static ssize_t node_read_meminfo(struct sys_device * dev,
 		       "Node %d WritebackTmp:   %8lu kB\n"
 		       "Node %d Slab:           %8lu kB\n"
 		       "Node %d SReclaimable:   %8lu kB\n"
-		       "Node %d SUnreclaim:     %8lu kB\n",
+		       "Node %d SUnreclaim:     %8lu kB\n"
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+		       "Node %d AnonHugePages:  %8lu kB\n"
+#endif
+			,
 		       nid, K(i.totalram),
 		       nid, K(i.freeram),
 		       nid, K(i.totalram - i.freeram),

@@ -122,7 +126,12 @@ static ssize_t node_read_meminfo(struct sys_device * dev,
 		       nid, K(node_page_state(nid, NR_WRITEBACK)),
 		       nid, K(node_page_state(nid, NR_FILE_PAGES)),
 		       nid, K(node_page_state(nid, NR_FILE_MAPPED)),
-		       nid, K(node_page_state(nid, NR_ANON_PAGES)),
+		       nid, K(node_page_state(nid, NR_ANON_PAGES)
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+			+ node_page_state(nid, NR_ANON_TRANSPARENT_HUGEPAGES) *
+			HPAGE_PMD_NR
+#endif
+		       ),
 		       nid, K(node_page_state(nid, NR_SHMEM)),
 		       nid, node_page_state(nid, NR_KERNEL_STACK) *
 				THREAD_SIZE / 1024,

@@ -133,7 +142,13 @@ static ssize_t node_read_meminfo(struct sys_device * dev,
 		       nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) +
 				node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
 		       nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)),
-		       nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
+		       nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE))
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+			, nid,
+			K(node_page_state(nid, NR_ANON_TRANSPARENT_HUGEPAGES) *
+				HPAGE_PMD_NR)
+#endif
+		       );
 	n += hugetlb_report_node_meminfo(nid, buf + n);
 	return n;
 }

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>