1Use multiple thread (de)compression in live migration
2=====================================================
3Copyright (C) 2015 Intel Corporation
4Author: Liang Li <liang.z.li@intel.com>
5
6This work is licensed under the terms of the GNU GPLv2 or later. See
7the COPYING file in the top-level directory.
8
9Contents:
10=========
11* Introduction
12* When to use
13* Performance
14* Usage
15* TODO
16
17Introduction
18============
19Instead of sending the guest memory directly, this solution will
20compress the RAM page before sending; after receiving, the data will
21be decompressed. Using compression in live migration can help
22to reduce the data transferred about 60%, this is very useful when the
23bandwidth is limited, and the total migration time can also be reduced
24about 70% in a typical case. In addition to this, the VM downtime can be
25reduced about 50%. The benefit depends on data's compressibility in VM.
26
27The process of compression will consume additional CPU cycles, and the
28extra CPU cycles will increase the migration time. On the other hand,
29the amount of data transferred will decrease; this factor can reduce
30the total migration time. If the process of the compression is quick
31enough, then the total migration time can be reduced, and multiple
32thread compression can be used to accelerate the compression process.
33
34The decompression speed of Zlib is at least 4 times as quick as
35compression, if the source and destination CPU have equal speed,
36keeping the compression thread count 4 times the decompression
37thread count can avoid resource waste.
38
39Compression level can be used to control the compression speed and the
40compression ratio. High compression ratio will take more time, level 0
41stands for no compression, level 1 stands for the best compression
42speed, and level 9 stands for the best compression ratio. Users can
43select a level number between 0 and 9.
44
45
46When to use the multiple thread compression in live migration
47=============================================================
48Compression of data will consume extra CPU cycles; so in a system with
49high overhead of CPU, avoid using this feature. When the network
50bandwidth is very limited and the CPU resource is adequate, use of
51multiple thread compression will be very helpful. If both the CPU and
52the network bandwidth are adequate, use of multiple thread compression
53can still help to reduce the migration time.
54
55Performance
56===========
57Test environment:
58
59CPU: Intel(R) Xeon(R) CPU E5-2680 0 @ 2.70GHz
60Socket Count: 2
61RAM: 128G
62NIC: Intel I350 (10/100/1000Mbps)
63Host OS: CentOS 7 64-bit
64Guest OS: RHEL 6.5 64-bit
65Parameter: qemu-system-x86_64 -enable-kvm -smp 4 -m 4096
66 /share/ia32e_rhel6u5.qcow -monitor stdio
67
68There is no additional application is running on the guest when doing
69the test.
70
71
72Speed limit: 1000Gb/s
73---------------------------------------------------------------
74                    | original  | compress thread: 8
75                    |   way     | decompress thread: 2
76                    |           | compression level: 1
77---------------------------------------------------------------
78total time(msec):   |   3333    |  1833
79---------------------------------------------------------------
80downtime(msec):     |    100    |   27
81---------------------------------------------------------------
82transferred ram(kB):|  363536   | 107819
83---------------------------------------------------------------
84throughput(mbps):   |  893.73   | 482.22
85---------------------------------------------------------------
86total ram(kB):      |  4211524  | 4211524
87---------------------------------------------------------------
88
89There is an application running on the guest which write random numbers
90to RAM block areas periodically.
91
92Speed limit: 1000Gb/s
93---------------------------------------------------------------
94                    | original  | compress thread: 8
95                    |   way     | decompress thread: 2
96                    |           | compression level: 1
97---------------------------------------------------------------
98total time(msec):   |   37369   | 15989
99---------------------------------------------------------------
100downtime(msec):     |    337    |  173
101---------------------------------------------------------------
102transferred ram(kB):|  4274143  | 1699824
103---------------------------------------------------------------
104throughput(mbps):   |  936.99   | 870.95
105---------------------------------------------------------------
106total ram(kB):      |  4211524  | 4211524
107---------------------------------------------------------------
108
109Usage
110=====
1111. Verify both the source and destination QEMU are able
112to support the multiple thread compression migration:
113    {qemu} info_migrate_capabilities
114    {qemu} ... compress: off ...
115
1162. Activate compression on the source:
117    {qemu} migrate_set_capability compress on
118
1193. Set the compression thread count on source:
120    {qemu} migrate_set_parameter compress_threads 12
121
1224. Set the compression level on the source:
123    {qemu} migrate_set_parameter compress_level 1
124
1255. Set the decompression thread count on destination:
126    {qemu} migrate_set_parameter decompress_threads 3
127
1286. Start outgoing migration:
129    {qemu} migrate -d tcp:destination.host:4444
130    {qemu} info migrate
131    Capabilities: ... compress: on
132    ...
133
134The following are the default settings:
135    compress: off
136    compress_threads: 8
137    decompress_threads: 2
138    compress_level: 1 (which means best speed)
139
140So, only the first two steps are required to use the multiple
141thread compression in migration. You can do more if the default
142settings are not appropriate.
143
144TODO
145====
146Some faster (de)compression method such as LZ4 and Quicklz can help
147to reduce the CPU consumption when doing (de)compression. If using
148these faster (de)compression method, less (de)compression threads
149are needed when doing the migration.
150