zfs/debian/patches/0004-increase-default-zcmd-allocation-to-256K.patch

   1 From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
   2 From: Colin Ian King <colin.king@canonical.com>
   3 Date: Tue, 31 Oct 2017 19:12:42 +0800
   4 Subject: [PATCH] increase-default-zcmd-allocation-to-256K
   5
   6 Increase default zcmd allocation to 256K (LP: #567557)
   7
   8 When creating hundreds of clones (for example using containers with
   9 LXD) cloning slows down as the number of clones increases over time.
  10 The reason for this is that the fetching of the clone information
  11 using a small zcmd buffer requires two ioctl calls, one to determine
  12 the size and a second to return the data. However, this requires
  13 gathering the data twice, once to determine the size and again to
  14 populate the zcmd buffer to return it to userspace.
  15
  16 These are expensive ioctl() calls, so instead, make the default buffer
  17 size much larger: 256K. This may sound large, but on 64 bit systems
  18 running ZFS this is not a huge chunk of memory for the speed
  19 improvement we gains for large sets of clones:
  20
  21         16K zcmd        256K zcmd
  22 Clones  Time    Clones  Time    Clone   % improvement
  23         (secs)  per sec (secs)  per sec
  24 100     7       14.29   5       20.00   28.57
  25 200     10      20.00   9       22.22   10.00
  26 300     19      15.79   18      16.67   5.26
  27 400     22      18.18   22      18.18   0.00
  28 500     29      17.24   29      17.24   0.00
  29 600     39      15.38   39      15.38   0.00
  30 700     46      15.22   45      15.56   2.17
  31 800     58      13.79   51      15.69   12.07
  32 900     74      12.16   61      14.75   17.57
  33 1000    90      11.11   74      13.51   17.78
  34 1100    98      11.22   87      12.64   11.22
  35 1200    102     11.76   95      12.63   6.86
  36 1300    113     11.50   104     12.50   7.96
  37 1400    143     9.79    109     12.84   23.78
  38 1500    145     10.34   132     11.36   8.97
  39 1600    165     9.70    145     11.03   12.12
  40 1700    187     9.09    156     10.90   16.58
  41 1800    210     8.57    166     10.84   20.95
  42 1900    226     8.41    183     10.38   19.03
  43 2000    256     7.81    198     10.10   22.66
  44 2200    311     7.07    238     9.24    23.47
  45 2400    373     6.43    271     8.86    27.35
  46 2600    487     5.34    316     8.23    35.11
  47 3000    619     4.85    426     7.04    31.18
  48 3400    915     3.72    549     6.19    40.00
  49 4000    1332    3.00    923     4.33    30.71
  50
  51 As one can see, with > 2000 clones we get 25-40% speed
  52 improvement.
  53
  54 This patch was originally suggested by Brian Behlendorf
  55 (see https://github.com/zfsonlinux/zfs/issues/6372), however
  56 this fix is a more generic fix to cover all zcmd cases.
  57
  58 Signed-off-by: Colin Ian King <colin.king@canonical.com>
  59 ---
  60  lib/libzfs/libzfs_util.c | 2 +-
  61  1 file changed, 1 insertion(+), 1 deletion(-)
  62
  63 diff --git a/lib/libzfs/libzfs_util.c b/lib/libzfs/libzfs_util.c
  64 index 19bb57ad4..2a069ffe6 100644
  65 --- a/lib/libzfs/libzfs_util.c
  66 +++ b/lib/libzfs/libzfs_util.c
  67 @@ -1134,7 +1134,7 @@ int
  68  zcmd_alloc_dst_nvlist(libzfs_handle_t *hdl, zfs_cmd_t *zc, size_t len)
  69  {
  70         if (len == 0)
  71 -               len = 16 * 1024;
  72 +               len = 256 * 1024;
  73         zc->zc_nvlist_dst_size = len;
  74         zc->zc_nvlist_dst =
  75             (uint64_t)(uintptr_t)zfs_alloc(hdl, zc->zc_nvlist_dst_size);