Root/Documentation/block/cfq-iosched.txt

1CFQ ioscheduler tunables
2========================
3
4slice_idle
5----------
6This specifies how long CFQ should idle for next request on certain cfq queues
7(for sequential workloads) and service trees (for random workloads) before
8queue is expired and CFQ selects next queue to dispatch from.
9
10By default slice_idle is a non-zero value. That means by default we idle on
11queues/service trees. This can be very helpful on highly seeky media like
12single spindle SATA/SAS disks where we can cut down on overall number of
13seeks and see improved throughput.
14
15Setting slice_idle to 0 will remove all the idling on queues/service tree
16level and one should see an overall improved throughput on faster storage
17devices like multiple SATA/SAS disks in hardware RAID configuration. The down
18side is that isolation provided from WRITES also goes down and notion of
19IO priority becomes weaker.
20
21So depending on storage and workload, it might be useful to set slice_idle=0.
22In general I think for SATA/SAS disks and software RAID of SATA/SAS disks
23keeping slice_idle enabled should be useful. For any configurations where
24there are multiple spindles behind single LUN (Host based hardware RAID
25controller or for storage arrays), setting slice_idle=0 might end up in better
26throughput and acceptable latencies.
27
28CFQ IOPS Mode for group scheduling
29===================================
30Basic CFQ design is to provide priority based time slices. Higher priority
31process gets bigger time slice and lower priority process gets smaller time
32slice. Measuring time becomes harder if storage is fast and supports NCQ and
33it would be better to dispatch multiple requests from multiple cfq queues in
34request queue at a time. In such scenario, it is not possible to measure time
35consumed by single queue accurately.
36
37What is possible though is to measure number of requests dispatched from a
38single queue and also allow dispatch from multiple cfq queue at the same time.
39This effectively becomes the fairness in terms of IOPS (IO operations per
40second).
41
42If one sets slice_idle=0 and if storage supports NCQ, CFQ internally switches
43to IOPS mode and starts providing fairness in terms of number of requests
44dispatched. Note that this mode switching takes effect only for group
45scheduling. For non-cgroup users nothing should change.
46
47CFQ IO scheduler Idling Theory
48===============================
49Idling on a queue is primarily about waiting for the next request to come
50on same queue after completion of a request. In this process CFQ will not
51dispatch requests from other cfq queues even if requests are pending there.
52
53The rationale behind idling is that it can cut down on number of seeks
54on rotational media. For example, if a process is doing dependent
55sequential reads (next read will come on only after completion of previous
56one), then not dispatching request from other queue should help as we
57did not move the disk head and kept on dispatching sequential IO from
58one queue.
59
60CFQ has following service trees and various queues are put on these trees.
61
62    sync-idle sync-noidle async
63
64All cfq queues doing synchronous sequential IO go on to sync-idle tree.
65On this tree we idle on each queue individually.
66
67All synchronous non-sequential queues go on sync-noidle tree. Also any
68request which are marked with REQ_NOIDLE go on this service tree. On this
69tree we do not idle on individual queues instead idle on the whole group
70of queues or the tree. So if there are 4 queues waiting for IO to dispatch
71we will idle only once last queue has dispatched the IO and there is
72no more IO on this service tree.
73
74All async writes go on async service tree. There is no idling on async
75queues.
76
77CFQ has some optimizations for SSDs and if it detects a non-rotational
78media which can support higher queue depth (multiple requests at in
79flight at a time), then it cuts down on idling of individual queues and
80all the queues move to sync-noidle tree and only tree idle remains. This
81tree idling provides isolation with buffered write queues on async tree.
82
83FAQ
84===
85Q1. Why to idle at all on queues marked with REQ_NOIDLE.
86
87A1. We only do tree idle (all queues on sync-noidle tree) on queues marked
88    with REQ_NOIDLE. This helps in providing isolation with all the sync-idle
89    queues. Otherwise in presence of many sequential readers, other
90    synchronous IO might not get fair share of disk.
91
92    For example, if there are 10 sequential readers doing IO and they get
93    100ms each. If a REQ_NOIDLE request comes in, it will be scheduled
94    roughly after 1 second. If after completion of REQ_NOIDLE request we
95    do not idle, and after a couple of milli seconds a another REQ_NOIDLE
96    request comes in, again it will be scheduled after 1second. Repeat it
97    and notice how a workload can lose its disk share and suffer due to
98    multiple sequential readers.
99
100    fsync can generate dependent IO where bunch of data is written in the
101    context of fsync, and later some journaling data is written. Journaling
102    data comes in only after fsync has finished its IO (atleast for ext4
103    that seemed to be the case). Now if one decides not to idle on fsync
104    thread due to REQ_NOIDLE, then next journaling write will not get
105    scheduled for another second. A process doing small fsync, will suffer
106    badly in presence of multiple sequential readers.
107
108    Hence doing tree idling on threads using REQ_NOIDLE flag on requests
109    provides isolation from multiple sequential readers and at the same
110    time we do not idle on individual threads.
111
112Q2. When to specify REQ_NOIDLE
113A2. I would think whenever one is doing synchronous write and not expecting
114    more writes to be dispatched from same context soon, should be able
115    to specify REQ_NOIDLE on writes and that probably should work well for
116    most of the cases.
117

Archive Download this file



interactive