as-iosched.txt

Anticipatory IO scheduler
-------------------------
Nick Piggin <piggin@cyberone.com.au>    13 Sep 2003

Attention! Database servers, especially those using "TCQ" disks should
investigate performance with the 'deadline' IO scheduler. Any system with high
disk performance requirements should do so, in fact.

If you see unusual performance characteristics of your disk systems, or you
see big performance regressions versus the deadline scheduler, please email
me. Database users don't bother unless you're willing to test a lot of patches
from me ;) its a known issue.

Also, users with hardware RAID controllers, doing striping, may find
highly variable performance results with using the as-iosched. The
as-iosched anticipatory implementation is based on the notion that a disk
device has only one physical seeking head.  A striped RAID controller
actually has a head for each physical device in the logical RAID device.

However, setting the antic_expire (see tunable parameters below) produces
very similar behavior to the deadline IO scheduler.


Selecting IO schedulers
-----------------------
To choose IO schedulers at boot time, use the argument 'elevator=deadline'.
'noop' and 'as' (the default) are also available. IO schedulers are assigned
globally at boot time only presently.


Anticipatory IO scheduler Policies
----------------------------------
The as-iosched implementation implements several layers of policies
to determine when an IO request is dispatched to the disk controller.
Here are the policies outlined, in order of application.

1. one-way Elevator algorithm.

The elevator algorithm is similar to that used in deadline scheduler, with
the addition that it allows limited backward movement of the elevator
(i.e. seeks backwards).  A seek backwards can occur when choosing between
two IO requests where one is behind the elevator's current position, and
the other is in front of the elevator's position. If the seek distance to
the request in back of the elevator is less than half the seek distance to
the request in front of the elevator, then the request in back can be chosen.
Backward seeks are also limited to a maximum of MAXBACK (1024*1024) sectors.
This favors forward movement of the elevator, while allowing opportunistic
"short" backward seeks.

2. FIFO expiration times for reads and for writes.

This is again very similar to the deadline IO scheduler.  The expiration
times for requests on these lists is tunable using the parameters read_expire
and write_expire discussed below.  When a read or a write expires in this way,
the IO scheduler will interrupt its current elevator sweep or read anticipation
to service the expired request.

3. Read and write request batching

A batch is a collection of read requests or a collection of write
requests.  The as scheduler alternates dispatching read and write batches
to the driver.  In the case a read batch, the scheduler submits read
requests to the driver as long as there are read requests to submit, and
the read batch time limit has not been exceeded (read_batch_expire).
The read batch time limit begins counting down only when there are
competing write requests pending.

In the case of a write batch, the scheduler submits write requests to
the driver as long as there are write requests available, and the
write batch time limit has not been exceeded (write_batch_expire).
However, the length of write batches will be gradually shortened
when read batches frequently exceed their time limit.

When changing between batch types, the scheduler waits for all requests
from the previous batch to complete before scheduling requests for the
next batch.

The read and write fifo expiration times described in policy 2 above
are checked only when in scheduling IO of a batch for the corresponding
(read/write) type.  So for example, the read FIFO timeout values are
tested only during read batches.  Likewise, the write FIFO timeout
values are tested only during write batches.  For this reason,
it is generally not recommended for the read batch time
to be longer than the write expiration time, nor for the write batch
time to exceed the read expiration time (see tunable parameters below).

When the IO scheduler changes from a read to a write batch,
it begins the elevator from the request that is on the head of the
write expiration FIFO.  Likewise, when changing from a write batch to
a read batch, scheduler begins the elevator from the first entry
on the read expiration FIFO.

4. Read anticipation.

Read anticipation occurs only when scheduling a read batch.
This implementation of read anticipation allows only one read request
to be dispatched to the disk controller at a time.  In
contrast, many write requests may be dispatched to the disk controller
at a time during a write batch.  It is this characteristic that can make
the anticipatory scheduler perform anomalously with controllers supporting
TCQ, or with hardware striped RAID devices. Setting the antic_expire
queue paramter (see below) to zero disables this behavior, and the anticipatory
scheduler behaves essentially like the deadline scheduler.

When read anticipation is enabled (antic_expire is not zero), reads
are dispatched to the disk controller one at a time.
At the end of each read request, the IO scheduler examines its next
candidate read request from its sorted read list.  If that next request
is from the same process as the request that just completed,
or if the next request in the queue is "very close" to the
just completed request, it is dispatched immediately.  Otherwise,
statistics (average think time, average seek distance) on the process
that submitted the just completed request are examined.  If it seems
likely that that process will submit another request soon, and that
request is likely to be near the just completed request, then the IO
scheduler will stop dispatching more read requests for up time (antic_expire)
milliseconds, hoping that process will submit a new request near the one
that just completed.  If such a request is made, then it is dispatched
immediately.  If the antic_expire wait time expires, then the IO scheduler
will dispatch the next read request from the sorted read queue.

To decide whether an anticipatory wait is worthwhile, the scheduler
maintains statistics for each process that can be used to compute
mean "think time" (the time between read requests), and mean seek
distance for that process.  One observation is that these statistics
are associated with each process, but those statistics are not associated
with a specific IO device.  So for example, if a process is doing IO
on several file systems on separate devices, the statistics will be
a combination of IO behavior from all those devices.


Tuning the anticipatory IO scheduler
------------------------------------
When using 'as', the anticipatory IO scheduler there are 5 parameters under
/sys/block/*/queue/iosched/. All are units of milliseconds.

The parameters are:
* read_expire
    Controls how long until a read request becomes "expired". It also controls the
    interval between which expired requests are served, so set to 50, a request
    might take anywhere < 100ms to be serviced _if_ it is the next on the
    expired list. Obviously request expiration strategies won't make the disk
    go faster. The result basically equates to the timeslice a single reader
    gets in the presence of other IO. 100*((seek time / read_expire) + 1) is
    very roughly the % streaming read efficiency your disk should get with
    multiple readers.

* read_batch_expire
    Controls how much time a batch of reads is given before pending writes are
    served. A higher value is more efficient. This might be set below read_expire
    if writes are to be given higher priority than reads, but reads are to be
    as efficient as possible when there are no writes. Generally though, it
    should be some multiple of read_expire.

* write_expire, and
* write_batch_expire are equivalent to the above, for writes.

* antic_expire
    Controls the maximum amount of time we can anticipate a good read (one
    with a short seek distance from the most recently completed request) before
    giving up. Many other factors may cause anticipation to be stopped early,
    or some processes will not be "anticipated" at all. Should be a bit higher
    for big seek time devices though not a linear correspondence - most
    processes have only a few ms thinktime.