Re: [PATCHv8 00/32] THP-enabled tmpfs/shmem using compound pages

[neha agarwal <neha.agbk@xxxxxxxxx>]

Hi All,

I have been testing Hugh's and Kirill's huge tmpfs patch sets with Cassandra (NoSQL database). I am seeing significant performance gap between these two implementations (~30%). Hugh's implementation performs better than Kirill's implementation. I am surprised why I am seeing this performance gap. Following is my test setup.

Patchsets

========
- For Hugh's: 
I checked out 4.6-rc3, applied Hugh's preliminary patches (01 to 10 patches) from here: https://lkml.org/lkml/2016/4/5/792 and then applied the THP patches posted on April 16 (01 to 29 patches).

- For Kirill's: 
I am using his branch  "git://git.kernel.org/pub/scm/linux/kernel/git/kas/linux.git hugetmpfs/v8", which is based off of 4.6-rc3, posted on May 12.

Khugepaged settings

================
cd /sys/kernel/mm/transparent_hugepage
echo 10 >khugepaged/alloc_sleep_millisecs
echo 10 >khugepaged/scan_sleep_millisecs
echo 511 >khugepaged/max_ptes_none


Mount options

===========
- For Hugh's:
sudo sysctl -w vm/shmem_huge=2
sudo mount -o remount,huge=1 /hugetmpfs

- For Kirill's:
sudo mount -o remount,huge=always /hugetmpfs
echo force > /sys/kernel/mm/transparent_hugepage/shmem_enabled
echo 511 >khugepaged/max_ptes_swap


Workload Setting

=============
Please look at the attached setup document for Cassandra (NoSQL database): cassandra-setup.txt


Machine setup

===========
36-core (72 hardware thread) dual-socket x86 server with 512 GB RAM running Ubuntu. I use control groups for resource isolation. Server and client threads run on different sockets. Frequency governor set to "performance" to remove any performance fluctuations due to frequency variation.


Throughput numbers

================
Hugh's implementation: 74522.08 ops/sec
Kirill's implementation: 54919.10 ops/sec


I am not sure if something is fishy with my test environment or if there is actually a performance gap between the two implementations. I have run this test 5-6 times so I am certain that this experiment is repeatable. I will appreciate if someone can help me understand the reason for this performance gap.

On Thu, May 12, 2016 at 11:40 AM, Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx> wrote:

This update aimed to address my todo list from lsf/mm summit:

 - we now able to recovery memory by splitting huge pages partly beyond
   i_size. This should address concern about small files.

 - bunch of bug fixes for khugepaged, including fix for data corruption
   reported by Hugh.

 - Disabled for Power as it requires deposited page table to get THP
   mapped and we don't do deposit/withdraw for file THP.

The main part of patchset (up to khugepaged stuff) is relatively stable --
I fixed few minor bugs there, but nothing major.

I would appreciate rigorous review of khugepaged and code to split huge
pages under memory pressure.

The patchset is on top of v4.6-rc3 plus Hugh's "easy preliminaries to
THPagecache" and Ebru's khugepaged swapin patches form -mm tree.

Git tree:

git://git.kernel.org/pub/scm/linux/kernel/git/kas/linux.git hugetmpfs/v8

== Changelog ==

v8:
  - khugepaged updates:
    + mark collapsed page dirty, otherwise vmscan would discard it;
    + account pages to mapping->nrpages on shmem_charge;
    + fix a situation when not all tail pages put on radix tree on collapse;
    + fix off-by-one in loop-exit condition in khugepaged_scan_shmem();
    + use radix_tree_iter_next/radix_tree_iter_retry instead of gotos;
    + fix build withount CONFIG_SHMEM (again);
  - split huge pages beyond i_size under memory pressure;
  - disable huge tmpfs on Power, as it makes use of deposited page tables,
    we don't have;
  - fix filesystem size limit accouting;
  - mark page referenced on split_huge_pmd() if the pmd is young;
  - uncharge pages from shmem, removed during split_huge_page();
  - make shmem_inode_info::lock irq-safe -- required by khugepaged;

v7:
  - khugepaged updates:
    + fix page leak/page cache corruption on collapse fail;
    + filter out VMAs not suitable for huge pages due misaligned vm_pgoff;
    + fix build without CONFIG_SHMEM;
    + drop few over-protective checks;
  - fix bogus VM_BUG_ON() in __delete_from_page_cache();

v6:
  - experimental collapse support;
  - fix swapout mapped huge pages;
  - fix page leak in faularound code;
  - fix exessive huge page allocation with huge=within_size;
  - rename VM_NO_THP to VM_NO_KHUGEPAGED;
  - fix condition in hugepage_madvise();
  - accounting reworked again;

v5:
  - add FileHugeMapped to /proc/PID/smaps;
  - make FileHugeMapped in meminfo aligned with other fields;
  - Documentation/vm/transhuge.txt updated;

v4:
  - first four patch were applied to -mm tree;
  - drop pages beyond i_size on split_huge_pages;
  - few small random bugfixes;

v3:
  - huge= mountoption now can have values always, within_size, advice and
    never;
  - sysctl handle is replaced with sysfs knob;
  - MADV_HUGEPAGE/MADV_NOHUGEPAGE is now respected on page allocation via
    page fault;
  - mlock() handling had been fixed;
  - bunch of smaller bugfixes and cleanups.

== Design overview ==

Huge pages are allocated by shmem when it's allowed (by mount option) and
there's no entries for the range in radix-tree. Huge page is represented by
HPAGE_PMD_NR entries in radix-tree.

MM core maps a page with PMD if ->fault() returns huge page and the VMA is
suitable for huge pages (size, alignment). There's no need into two
requests to file system: filesystem returns huge page if it can,
graceful fallback to small pages otherwise.

As with DAX, split_huge_pmd() is implemented by unmapping the PMD: we can
re-fault the page with PTEs later.

Basic scheme for split_huge_page() is the same as for anon-THP.
Few differences:

  - File pages are on radix-tree, so we have head->_count offset by
    HPAGE_PMD_NR. The count got distributed to small pages during split.

  - mapping->tree_lock prevents non-lockless access to pages under split
    over radix-tree;

  - Lockless access is prevented by setting the head->_count to 0 during
    split, so get_page_unless_zero() would fail;

  - After split, some pages can be beyond i_size. We drop them from
    radix-tree.

  - We don't setup migration entries. Just unmap pages. It helps
    handling cases when i_size is in the middle of the page: no need
    handle unmap pages beyond i_size manually.

COW mapping handled on PTE-level. It's not clear how beneficial would be
allocation of huge pages on COW faults. And it would require some code to
make them work.

I think at some point we can consider teaching khugepaged to collapse
pages in COW mappings, but allocating huge on fault is probably overkill.

As with anon THP, we mlock file huge page only if it mapped with PMD.
PTE-mapped THPs are never mlocked. This way we can avoid all sorts of
scenarios when we can leak mlocked page.

As with anon THP, we split huge page on swap out.

Truncate and punch hole that only cover part of THP range is implemented
by zero out this part of THP.

This have visible effect on fallocate(FALLOC_FL_PUNCH_HOLE) behaviour.
As we don't really create hole in this case, lseek(SEEK_HOLE) may have
inconsistent results depending what pages happened to be allocated.
I don't think this will be a problem.

We track per-super_block list of inodes which potentially have huge page
partly beyond i_size. Under memory pressure or if we hit -ENOSPC, we split
such pages in order to recovery memory.

The list is per-sb, as we need to split a page from our filesystem if hit
-ENOSPC (-o size= limit) during shmem_getpage_gfp() to free some space.

Hugh Dickins (1):
  shmem: get_unmapped_area align huge page

Kirill A. Shutemov (31):
  thp, mlock: update unevictable-lru.txt
  mm: do not pass mm_struct into handle_mm_fault
  mm: introduce fault_env
  mm: postpone page table allocation until we have page to map
  rmap: support file thp
  mm: introduce do_set_pmd()
  thp, vmstats: add counters for huge file pages
  thp: support file pages in zap_huge_pmd()
  thp: handle file pages in split_huge_pmd()
  thp: handle file COW faults
  thp: skip file huge pmd on copy_huge_pmd()
  thp: prepare change_huge_pmd() for file thp
  thp: run vma_adjust_trans_huge() outside i_mmap_rwsem
  thp: file pages support for split_huge_page()
  thp, mlock: do not mlock PTE-mapped file huge pages
  vmscan: split file huge pages before paging them out
  page-flags: relax policy for PG_mappedtodisk and PG_reclaim
  radix-tree: implement radix_tree_maybe_preload_order()
  filemap: prepare find and delete operations for huge pages
  truncate: handle file thp
  mm, rmap: account shmem thp pages
  shmem: prepare huge= mount option and sysfs knob
  shmem: add huge pages support
  shmem, thp: respect MADV_{NO,}HUGEPAGE for file mappings
  thp: update Documentation/vm/transhuge.txt
  thp: extract khugepaged from mm/huge_memory.c
  khugepaged: move up_read(mmap_sem) out of khugepaged_alloc_page()
  shmem: make shmem_inode_info::lock irq-safe
  khugepaged: add support of collapse for tmpfs/shmem pages
  thp: introduce CONFIG_TRANSPARENT_HUGE_PAGECACHE
  shmem: split huge pages beyond i_size under memory pressure

 Documentation/filesystems/Locking    |   10 +-
 Documentation/vm/transhuge.txt       |  130 ++-
 Documentation/vm/unevictable-lru.txt |   21 +
 arch/alpha/mm/fault.c                |    2 +-
 arch/arc/mm/fault.c                  |    2 +-
 arch/arm/mm/fault.c                  |    2 +-
 arch/arm64/mm/fault.c                |    2 +-
 arch/avr32/mm/fault.c                |    2 +-
 arch/cris/mm/fault.c                 |    2 +-
 arch/frv/mm/fault.c                  |    2 +-
 arch/hexagon/mm/vm_fault.c           |    2 +-
 arch/ia64/mm/fault.c                 |    2 +-
 arch/m32r/mm/fault.c                 |    2 +-
 arch/m68k/mm/fault.c                 |    2 +-
 arch/metag/mm/fault.c                |    2 +-
 arch/microblaze/mm/fault.c           |    2 +-
 arch/mips/mm/fault.c                 |    2 +-
 arch/mn10300/mm/fault.c              |    2 +-
 arch/nios2/mm/fault.c                |    2 +-
 arch/openrisc/mm/fault.c             |    2 +-
 arch/parisc/mm/fault.c               |    2 +-
 arch/powerpc/mm/copro_fault.c        |    2 +-
 arch/powerpc/mm/fault.c              |    2 +-
 arch/s390/mm/fault.c                 |    2 +-
 arch/score/mm/fault.c                |    2 +-
 arch/sh/mm/fault.c                   |    2 +-
 arch/sparc/mm/fault_32.c             |    4 +-
 arch/sparc/mm/fault_64.c             |    2 +-
 arch/tile/mm/fault.c                 |    2 +-
 arch/um/kernel/trap.c                |    2 +-
 arch/unicore32/mm/fault.c            |    2 +-
 arch/x86/mm/fault.c                  |    2 +-
 arch/xtensa/mm/fault.c               |    2 +-
 drivers/base/node.c                  |   13 +-
 drivers/char/mem.c                   |   24 +
 drivers/iommu/amd_iommu_v2.c         |    3 +-
 drivers/iommu/intel-svm.c            |    2 +-
 fs/proc/meminfo.c                    |    7 +-
 fs/proc/task_mmu.c                   |   10 +-
 fs/userfaultfd.c                     |   22 +-
 include/linux/huge_mm.h              |   36 +-
 include/linux/khugepaged.h           |    6 +
 include/linux/mm.h                   |   51 +-
 include/linux/mmzone.h               |    4 +-
 include/linux/page-flags.h           |   19 +-
 include/linux/radix-tree.h           |    1 +
 include/linux/rmap.h                 |    2 +-
 include/linux/shmem_fs.h             |   45 +-
 include/linux/userfaultfd_k.h        |    8 +-
 include/linux/vm_event_item.h        |    7 +
 include/trace/events/huge_memory.h   |    3 +-
 ipc/shm.c                            |   10 +-
 lib/radix-tree.c                     |   68 +-
 mm/Kconfig                           |    8 +
 mm/Makefile                          |    2 +-
 mm/filemap.c                         |  226 ++--
 mm/gup.c                             |    7 +-
 mm/huge_memory.c                     | 2032 ++++++----------------------------
 mm/internal.h                        |    4 +-
 mm/khugepaged.c                      | 1851 +++++++++++++++++++++++++++++++
 mm/ksm.c                             |    5 +-
 mm/memory.c                          |  860 +++++++-------
 mm/mempolicy.c                       |    4 +-
 mm/migrate.c                         |    5 +-
 mm/mmap.c                            |   26 +-
 mm/nommu.c                           |    3 +-
 mm/page-writeback.c                  |    1 +
 mm/page_alloc.c                      |   21 +
 mm/rmap.c                            |   78 +-
 mm/shmem.c                           |  918 +++++++++++++--
 mm/swap.c                            |    2 +
 mm/truncate.c                        |   22 +-
 mm/util.c                            |    6 +
 mm/vmscan.c                          |    6 +
 mm/vmstat.c                          |    4 +
 75 files changed, 4240 insertions(+), 2415 deletions(-)
 create mode 100644 mm/khugepaged.c

--
2.8.1

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

--

Thanks and Regards,

Neha Agarwal

University of Michigan

1. Download and extract Cassandra
http://archive.apache.org/dist/cassandra/2.0.16/apache-cassandra-2.0.16-bin.tar.gz

Note that my test version is Cassandra-2.0.16.
We will denote the path to which the file is extracted as CASSANDRA_BIN

2. Setup environment for cassandra
mkdir -p run_cassandra/cassandra_conf/triggers

- Download cassandra-env.sh, cassandra.yaml, log4j-server.properties from my mail
attachement and then copy those files in run_cassandra/cassandra_conf
- Search for /home/nehaag/hugetmpfs in these files and change this to a local
directory mounted as tmpfs. Let’s say that is CASSANDRA_DATA.  A folder named
"cassandra" will be automatically created (For example:
CASSANDRA_DATA/cassandra) when running Cassandra.
- Please note that these scripts will need modifications if you use Cassandra
version other that 2.0.16

- Download create-ycsb-table.cql.j2 from my email attachment and copy it in
run_cassandra/

3. JAVA setup, get JRE: openjdk v1.7.0_101 (sudo apt-get install openjdk-7-jre
for Ubuntu)

4. Setup YCSB Load generator:
- Clone ycsb from: https://github.com/brianfrankcooper/YCSB.git. Let’s say this is
downloaded to YCSB_ROOT
- You need to have maven 3 installed (`sudo apt-get install maven’ in ubuntu)
- Create a script (say run-cassandra.sh) in run_cassandra as follows:

input_file=run_cassandra/create-ycsb-table.cql.j2
cassandra_cli=${CASSANDRA_BIN}/bin/cassandra-cli
host=”127.0.0.1” #Ip address of the machine running cassasndra server
$cassandra_cli -h $host --jmxport 7199 -f create-ycsb-table.cql
cd ${YCSB_ROOT}

# Load dataset
${YCSB_ROOT}/bin/ycsb -cp ${YCSB_ROOT}/cassandra/target/dependency/slf4j-simple-1.7.12.jar:${YCSB_ROOT}/cassandra/target/dependency/slf4j-simple-1.7.12.jar load cassandra-10 -p hosts=$host -threads  20 -p fieldcount=20 -p recordcount=5000000 -P ${YCSB_ROOT}/workloads/workloadb -s

# Run benchmark
${YCSB_ROOT}/bin/ycsb -cp ${YCSB_ROOT}/cassandra/target/dependency/slf4j-simple-1.7.12.jar:${YCSB_ROOT}/cassandra/target/dependency/slf4j-simple-1.7.12.jar run cassandra-10 -p hosts=$host -threads  20 -p fieldcount=20 -p operationcount=50000000 -p recordcount=5000000 -p readproportion=0.05 -p updateproportion=0.95 -P ${YCSB_ROOT}/workloads/workloadb -s

5. Run the cassandra server on host machine:
rm -r ${CASSANDRA_DATA}/cassandra && CASSANDRA_CONF=run_cassandra/cassandra_conf JRE_HOME=/usr/lib/jvm/java-7-openjdk-amd64/jre ${CASSANDRA_BIN}/bin/cassandra -f

6. Run load generator on same/some other machine:
./run-cassandra.sh

YCSB periodcally spits out the throughput and latency number
At the end overall throughput and latency will be printed out

Attachment: cassandra-env.sh
Description: Bourne shell script

Attachment: cassandra.yaml
Description: Binary data

Attachment: create-ycsb-table.cql.j2
Description: Binary data

Attachment: log4j-server.properties
Description: Binary data