blob: 9198a418f39c8a8a3b0773c211e32c84ac944694 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
|
#!/bin/bash
prog=survey.sh
. ../.script-config
LOG=/home/`whoami`/zpios-logs/`uname -r`/kpios-`date +%Y%m%d`/
mkdir -p ${LOG}
# Apply all tunings described below to generate some best case
# numbers for what is acheivable with some more elbow grease.
NAME="prefetch+zerocopy+checksum+pending1024+kmem"
echo "----------------------- ${NAME} ------------------------------"
./zpios.sh \
"" \
"zfs_prefetch_disable=1 zfs_vdev_max_pending=1024 zio_bulk_flags=0x100" \
"--zerocopy" \
${LOG}/${NAME}/ \
"${CMDDIR}/zfs/zfs set checksum=off lustre" | \
tee ${LOG}/${NAME}.txt
# Baseline number for an out of the box config with no manual tuning.
# Ideally, we will want things to be automatically tuned and for this
# number to approach the tweaked out results above.
NAME="baseline"
echo "----------------------- ${NAME} ------------------------------"
./zpios.sh \
"" \
"" \
"" \
${LOG}/${NAME}/ | \
tee ${LOG}/${NAME}.txt
# Disable ZFS's prefetching. For some reason still not clear to me
# current prefetching policy is quite bad for a random workload.
# Allow the algorithm to detect a random workload and not do anything
# may be the way to address this issue.
NAME="prefetch"
echo "----------------------- ${NAME} ------------------------------"
./zpios.sh \
"" \
"zfs_prefetch_disable=1" \
"" \
${LOG}/${NAME}/ | \
tee ${LOG}/${NAME}.txt
# As expected, simulating a zerocopy IO path improves performance
# by freeing up lots of CPU which is wasted move data between buffers.
NAME="zerocopy"
echo "----------------------- ${NAME} ------------------------------"
./zpios.sh \
"" \
"" \
"--zerocopy" \
${LOG}/${NAME}/ | \
tee ${LOG}/${NAME}.txt
# Disabling checksumming should show some (if small) improvement
# simply due to freeing up a modest amount of CPU.
NAME="checksum"
echo "----------------------- ${NAME} ------------------------------"
./zpios.sh \
"" \
"" \
"" \
${LOG}/${NAME}/ \
"${CMDDIR}/zfs/zfs set checksum=off lustre" | \
tee ${LOG}/${NAME}.txt
# Increasing the pending IO depth also seems to improve things likely
# at the expense of latency. This should be exported more because I'm
# seeing a much bigger impact there that I would have expected. There
# may be some low hanging fruit to be found here.
NAME="pending"
echo "----------------------- ${NAME} ------------------------------"
./zpios.sh \
"" \
"zfs_vdev_max_pending=1024" \
"" \
${LOG}/${NAME}/ | \
tee ${LOG}/${NAME}.txt
# To avoid memory fragmentation issues our slab implementation can be
# based on a virtual address space. Interestingly, we take a pretty
# substantial performance penalty for this somewhere in the low level
# IO drivers. If we back the slab with kmem pages we see far better
# read performance numbers at the cost of memory fragmention and general
# system instability due to large allocations. This may be because of
# an optimization in the low level drivers due to the contigeous kmem
# based memory. This needs to be explained. The good news here is that
# with zerocopy interfaces added at the DMU layer we could gaurentee
# kmem based memory for a pool of pages.
#
# 0x100 = KMC_KMEM - Force kmem_* based slab
# 0x200 = KMC_VMEM - Force vmem_* based slab
NAME="kmem"
echo "----------------------- ${NAME} ------------------------------"
./zpios.sh \
"" \
"zio_bulk_flags=0x100" \
"" \
${LOG}/${NAME}/ | \
tee ${LOG}/${NAME}.txt
|
