[PATCH v2 00/12] multiblock allocator improvements

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



** Changed since v1 [2] **

 1. Rebase over Kemeng's recent mballoc patchset [3]
 2. Picked up Kemeng's RVB on patch 1/12

 [2] https://lore.kernel.org/all/cover.1685009579.git.ojaswin@xxxxxxxxxxxxx/
 [3] https://lore.kernel.org/all/20230417110617.2664129-1-shikemeng@xxxxxxxxxxxxxxx/

** Changes since RFC [1] **

[1] https://lore.kernel.org/linux-ext4/cover.1674822311.git.ojaswin@xxxxxxxxxxxxx/

1. Patch 1 reverts the commit 32c08693 

Lore link:
https://lore.kernel.org/linux-ext4/20230209194825.511043-15-shikemeng@xxxxxxxxxxxxxxx/ 

So this patch was intended to remove a dead if-condition but it was not
actually dead code and removing it was causing a performance regression.
Unfortunately I somehow missed that when I was reviewing his patchset
and it already went in so I had to revert the commit. I've added details
of the regression and root cause in the revert commit. Also attaching
the performance numbers I observer:

Workload: fsmark - 100GiB ramdisk, 64 threads writing ~42000 files nodelalloc
-----
Baseline kernel:             ~5000 files/sec, ~9,000,000 extents scanned

This patchset rebased on 
ted/dev w/o revert patch:    ~8000 files/sec, ~7,000,000 ex scanned (+40-50%)

This patchset on ted/dev 
with revert patch:           ~30000 files/sec, ~800,000 ex scanned (+500%)
-----

2. Added Patch 13 which introduces symbolic names for criterias

3. In CR1.5 patch (Patch 12), in ext4_mb_choose_next_group_cr1_5(),
while trimming we also consider stripe size now. If specified, we round
up the goal length to stripe size. Here, with bigalloc I've made an
assumption that stripe size in fs blocks is always a multiple of
cluster_ratio. This assumption is based on a yet unmerged patch:
https://lore.kernel.org/linux-ext4/20230417110617.2664129-5-shikemeng@xxxxxxxxxxxxxxx/

4. In CR1.5 patch, slight optimization in ext4_mb_choose_next_group_cr1_5()
based on Jan's feedback.

I've run xfstests quick on the patchset and plan to run auto overnight.
Would report if anything is breaking.

** Original Cover letter **

This patchset intends to improve some of the shortcomings of mb allocator
that we had noticed while running various tests and workloads in a
POWERPC machine with 64k block size.  

** Problems **

More specifically, we were seeing a sharp drop in performance when the
FS was highly fragmented (64K bs). We noticed that:

Problem 1: prefetch logic seemed to be skipping BLOCK_UNINIT groups
which was resulting in buddy and CR0/1 cache not being initialized for
these even though it could be done without any IO. (Not sure if there
was any history behind this design, do let me know if so).

Problem 2: With a 64K bs FS, we were commonly seeing cases where CR1
would correctly identify a good group but due to very high
fragmentation, complex scan would exit early due to ac->ac_found >
s_mb_max_to_scan, resulting in trimming of the allocated len.

Problem 3: Even though our avg free extent was say 4MB and original
request was merely 1 block of data, mballoc noramlization kept adding
PAs and requesting 8MB chunks. This led to almost all the requests
falling into slower CR 2 and with increased threads, we started seeing
lots of CR3 requests as well.

** How did we address them **

Problem 1 (Patch 8,9): Make ext4_mb_prefetch also call
ext4_read_block_bitmap_nowait() in case of BLOCK_UNINIT, so it can init
the BG and exit early without an IO. Next, fix the calls to
prefetch_fini so these newly init BGs can have their buddy initialised.

Problem 2 (Patch 7): When we come to complex_scan after CR1, my
understanding is that due to free/frag > goal_len, we can be sure that
there is atleast one chunk big enough to accomodate the goal request.
Hence, we can skip the overhead of mb_find_extent() other accounting for
each free extent and just process extents that are big enough.

Problem 3 (Patch 11): To solve this problem, this patchset implements a
new allocation criteria (CR1.5 or CR1_5 in code). The idea is that if
CR1 fails to find a BG, it will jump to CR1.5. Here the flow is as
follows:

  * We make an assumption that if CR1 has failed that means none of the
    currently cached BGs have a big enough continuous extent to satisfy
    our request In this case we fall to CR1.5.

  * In CR 1.5, we find the highest available free/frag BGs (from CR1
    lists) and trim the PAs to this order so that we can find 
    a BG without IO overhead of CR2. 
    
  * Parallely, prefetch will get in more groups in memory, and as more
    and more groups are cached, CR1.5 becomes a better replacement of
    CR2. This is because, for example, if all BGs are cahced and we
    couldn't find anything in CR0/1, we can assume that no BG has a big
    enough continuous free extent and hence CR1.5 can directly trim and
    find the next biggest extent we could get. In this scenario, without
    CR1.5, we would have continued scanning in CR2 which would have
    most probably trimmed the request after scanning for ~200 extents.
    
CR1.5 results in improved allocation speed at the cost of slightly increased
trimming of the len of blocks allocated. 

** Performance Numbers **

Unless stated otherwise, these numbers are from fsmark and fio tests with 64k
BS, 64K pagesize on 100Gi nvme0n1 with nodelalloc. There tests were performed
after the FS was fragmented till Avg Fragment Size == 4MB.

* Test 1: Writing ~40000 files of 64K each in a single directory (64 threads, fsmark)
* Test 2: Same as Test 1 on a 500GiB pmem device with dax
* Test 3: 5Gi write with mix of random and seq writes (fio)
* Test 4: 5Gi sequential writes (fio)

Here:
e = extents scanned
c = cr0 / cr1 / cr1.5 / cr2 / cr3 hits

+─────────+───────────────────────────────────+────────────────────────────────+
|         | Unpatched                         | Patched                        |
+─────────+───────────────────────────────────+────────────────────────────────+
| Test 1  | 6866 files/s                      | 13527 files/s                  |
|         | e: 8,188,644                      | e: 1,719,725                   |
|         | c: 381 / 330 / - / 4779 / 35534   | c: 381/ 280 / 33299/ 1000/ 6064|
+─────────+───────────────────────────────────+────────────────────────────────+
| Test 2  | 6927 files/s                      | 8422 files/s                   |
|         | e: 8,055,911                      | e: 261,268                     |
|         | cr: 1011 / 999 / - / 6153 / 32861 | c: 1721 / 1210 / 38093 / 0 / 0 |
+─────────+───────────────────────────────────+────────────────────────────────+
| Test 3  | 387 MiB/s                         | 443 MiB/s                      |
+─────────+───────────────────────────────────+────────────────────────────────+
| Test 4  | 3139 MiB/s                        | 3180 MiB/s                     |
+─────────+───────────────────────────────────+────────────────────────────────+

The numbers of same tests with 4k bs 64k pagesize are:

+─────────+────────────────────────────────────+────────────────────────────────+
|         | Unpatched                          | Patched                        |
+─────────+────────────────────────────────────+────────────────────────────────+
| Test 1  | 21618 files/s                      | 23528 files/s                  |
|         | e: 8,149,272                       | e: 223,013                     |
|         | c: 34 / 1380 / - / 5624 / 34710    | 34 / 1341 / 40387 / 0 / 0      |
+─────────+───────────────────────────────────+─────────────────────────────────+
| Test 2  | 30739 files/s                      | 30946 files/s                  |
|         | e: 7,742,853                       | e: 2,176,475                   |
|         | c: 1131 / 2244 / - / 3914 / 34468  | c: 1596/1079/28425/1098/8547   |
+─────────+───────────────────────────────────+─────────────────────────────────+
| Test 3  | 200 MiB/s                          | 186MiB/s                       |
+─────────+───────────────────────────────────+─────────────────────────────────+
| Test 4  | 621 MiB/s                          | 632 MiB/s                      |
+─────────+────────────────────────────────────+────────────────────────────────+

** Some Observations **

1. In the case of highly fragmented 64k blocksize most of the performance is
lost since we hold the BG lock while scanning a block group for best extent.
As our goal len is 8MB and we only have 4MB blocks, we are taking a long time
to scan causing other threads to wait on the BG lock. This can be seen in perf
diff of unpatched vs patched:

    83.14%    -24.89%  [kernel.vmlinux]            [k] do_raw_spin_lock

Using lockstat and perf call graph I was able to confirm that this lock was the
BG lock taken in ext4_mb_regular_allocator, contending with other processes trying
to take the same BG's lock in ext4_mb_regular_allocator() and __ext4_new_inode()


2. Currently, I do see some increase in fragmentation, I can take this
up as future work. Below are the e2freefrag results after Test 1 with
64k BS:

xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

Unpatched:

Min. free extent: 128 KB
Max. free extent: 8000 KB
Avg. free extent: 4096 KB
Num. free extent: 12630

HISTOGRAM OF FREE EXTENT SIZES:
Extent Size Range :  Free extents   Free Blocks  Percent
  128K...  256K-  :             1             2    0.00%
  256K...  512K-  :             1             6    0.00%
  512K... 1024K-  :             4            48    0.01%
    1M...    2M-  :             5           120    0.01%
    2M...    4M-  :         11947        725624   85.31%
    4M...    8M-  :           672         83796    9.85%

Patched:

Min. free extent: 64 KB
Max. free extent: 11648 KB
Avg. free extent: 2688 KB
Num. free extent: 18847

HISTOGRAM OF FREE EXTENT SIZES:
Extent Size Range :  Free extents   Free Blocks  Percent
   64K...  128K-  :             1             1    0.00%
  128K...  256K-  :             2             5    0.00%
  256K...  512K-  :             1             5    0.00%
  512K... 1024K-  :           297          3909    0.48%
    1M...    2M-  :         11221        341065   42.13%
    2M...    4M-  :          4940        294260   36.35%
    4M...    8M-  :          2384        170169   21.02%
    8M...   16M-  :             1           182    0.02%

xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

-------------------------------------

Since these changes are looking good to me from my end, so posting for a
feedback from ext4 community.  

(gcexfstests -c all quick went fine with no new failures reported)

Any thoughts/suggestions are welcome!! 

Regards,
Ojaswin

Ojaswin Mujoo (10):
  Revert "ext4: remove ac->ac_found > sbi->s_mb_min_to_scan dead check
    in ext4_mb_check_limits"
  ext4: Convert mballoc cr (criteria) to enum
  ext4: Add per CR extent scanned counter
  ext4: Add counter to track successful allocation of goal length
  ext4: Avoid scanning smaller extents in BG during CR1
  ext4: Don't skip prefetching BLOCK_UNINIT groups
  ext4: Ensure ext4_mb_prefetch_fini() is called for all prefetched BGs
  ext4: Abstract out logic to search average fragment list
  ext4: Add allocation criteria 1.5 (CR1_5)
  ext4: Give symbolic names to mballoc criterias

Ritesh Harjani (IBM) (2):
  ext4: mballoc: Remove useless setting of ac_criteria
  ext4: Remove unused extern variables declaration

 fs/ext4/ext4.h              |  70 +++++-
 fs/ext4/mballoc.c           | 453 ++++++++++++++++++++++++++----------
 fs/ext4/mballoc.h           |  16 +-
 fs/ext4/super.c             |  11 +-
 fs/ext4/sysfs.c             |   2 +
 include/trace/events/ext4.h |  18 +-
 6 files changed, 428 insertions(+), 142 deletions(-)

-- 
2.31.1




[Index of Archives]     [Linux Ext4 Filesystem]     [Union Filesystem]     [Filesystem Testing]     [Ceph Users]     [Ecryptfs]     [NTFS 3]     [AutoFS]     [Kernel Newbies]     [Share Photos]     [Security]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux Cachefs]     [Reiser Filesystem]     [Linux RAID]     [NTFS 3]     [Samba]     [Device Mapper]     [CEPH Development]

  Powered by Linux