Electrical computers and digital data processing systems: input/ – Input/output data processing – Input/output access regulation
Reexamination Certificate
1999-01-29
2001-09-18
Shin, Christopher B. (Department: 2782)
Electrical computers and digital data processing systems: input/
Input/output data processing
Input/output access regulation
C710S005000, C710S052000, C710S263000, C709S241000
Reexamination Certificate
active
06292856
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Technical Field of the Invention
This invention relates to scheduling of input/output functions relative to input/out (I/O) devices in a multitasking data processing system. More particularly, it relates to a system and method for enabling applications to influence the order of service of an I/O request after the I/O request has been submitted.
2. Background Art
In contemporary multitasking data processing systems, such as is illustrated in
FIG. 1
, the operating system
32
of processor
20
schedules input/output operations with respect to input/output device
24
through device interface
34
over I/O bus
28
responsive to I/O requests from applications
30
. Typically, I/O device
28
is a long term storage device, such as a magnetic disk. Main storage
22
includes program and data storage which is accessible to processor
20
over bus
26
.
Referring to
FIG. 2
, in one contemporary multitasking data processing system, an operating system
32
schedules input/output tasks or operations in a first-come first-served manner. In such systems, if blocks A, B and C are written to disk, they would complete in that specific order. A user application
30
, which may or may not be running on the same system (for example, the file system could be a distributed file system), makes file access or update requests to the file system
40
. The file system then determines the disk I/O needed to satisfy the file requests made by user application
30
and issues I/O requests to the I/O driver
42
. The I/O driver formats the requests to a format required by the Media Manager
44
and sends the requests on to Media Manager
44
for processing. The Media Manager
44
performs no optimization or I/O ordering for efficiency, it handles requests first-come first-served. The Media Manager
44
creates the channel program
52
and sends the channel request to the device driver
46
which sends the request to the device
24
via the channel
28
. The I/O driver
42
represents I/O requests by a control block
54
that can later be waited on by units of work (hereafter referred to as tasks) that are processing the file requests from the application. All I/O performed in the system is asynchronous, that is, I/O requests are made to the I/O driver
42
and must later be waited on to ensure completion via another (WAIT) request to the I/O driver
42
and specifying the corresponding I/O control block
54
.
In this system, which is a logging file system, in order to reduce I/O rates in general, a group of blocks of data are read/written with respect to I/O device
24
at a time, rather than one physical disk block at a time. Metadata updates are logged to a logfile on disk, and metadata is only written to disk at periodic sync intervals and when the logfile is getting full so committed transactions can be freed and space made available in the logfile. Referring to
FIG. 3
, file system layer
40
uses five buffers in memory
22
to contain most recent written logfile pages
65
-
69
. When all five buffers
65
-
69
are I/O pending, file system
40
waits for first log file page
65
to complete so it can re-use memory
22
buffer for a different logfile page. When writing large files, the disk I/O access pattern illustrated in
FIG. 3
occurs repeatedly. That is, when application layer
30
writes large (over 1M) files to file system layer
40
, file system layer
40
buffers and writes user data in 64K increments
61
-
64
, etc. Thus, a 1M file entails 16 I/O requests
50
for user data, and would repeat the access pattern
61
-
65
four times. As logfile pages
65
-
68
fill, they have to be written out to disk
24
. Thus, writing a large file usually mans writing some user data
61
-
64
, and then writing a logfile page
65
. Since, in this specific system, there are five buffers
65
-
69
for logfile information, the first five instances of pattern
61
-
65
occur really fast. The first 1M file (which is four instances of pattern
61
-
65
, involving log files
65
-
68
) completes with no I/O waits and gets over lOM per second response time. The next 1M file, however, runs into the problem that the five buffers for the logfile I/O are pending I/O completion, and must wait for at least one of these I/Os to complete before proceeding. Thus, extra I/O wait time is incurred because the I/O requests for four data blocks
61
-
64
, and one logfile block
65
, must complete when writing the second file, and this repeats every 256K that is written to the second file. Note that when writing the second file, the user task is being made to wait for the I/O to complete for the first file (unnecessarily). This pattern repeats for subsequent files being written. This results in performance equivalent to synchronous file write performance instead of the asynchronous file performance the first file received. Any file after the first will get this synchronous write performance until it comes time to free up space in the log file
60
by writing metadata or the sync daemon (which writes metadata) runs. The default sync daemon runs every 30 seconds. When this happens, metadata write is added to the I/O queue in 256K increments and file system
40
waits for this I/O to complete before it can proceed. When log
60
is full, or a sync operation occurs, and since I/O is first-come, first-served, a wait for all preceding I/Os that are not completed occurs, and performance for the file being written goes way down to less that 300 KBs/sec.
The performance of I/O writes is then very erratic based on where the system is in the cycle when application
30
submits file write requests. Performance may be very fast because log buffers are available; sync rate because the logfile buffer must be made available before it can be written; very slow because metadata is being written out and all I/O needs to be complete before continuing; or very fast because as soon as the metadata is written, the system has all logfile buffers available to it.
In some contemporary multitasking data processing systems, operating systems execute various approaches for scheduling input/output requests or operations in sequences based upon task priorities.
For example, in one such system, U.S. Pat. No. 5,220,653 by F. Miro for scheduling input/output operations in multitasking systems, of common assignee, the teachings of which are incorporated herein by this reference, an improvement on elevator I/O scheduling is described. Elevator I/O scheduling is scheduling based on which part of the disk the I/O will read or write. The Miro priority scheme allows a priority to be associated with the I/O request when it is scheduled and that priority is used to order the I/O in importance. Miro also provides algorithms for dealing with starvation. Thus, Miro provides for sorting I/Os based on priority in such a manner as to prevent starvation. However, Miro does not provide a method or system for allowing an application to adjust I/O priority after the application has submitted to I/O request. Because an application does not necessarily know at the time of submission the priority required by its I/O request request, there is a need in the art for a method and system which allows the application to adjust the priority or even cancel the I/O request after its submission.
Generally, contemporary systems require that a user submit priority indicia concurrently with an I/O request, and then is at the mercy of the I/O code which handles the request to do any performance optimization, such as writing out data that is closest to he current device head next, or write out the smallest file next, and so forth. In such systems, the user application has no control or influence over its I/O requests after the I/O request is submitted.
The above problem and problems like it cannot be solved by the prior art—either using the elevator disk scheduling technique and/or the priority sorting based on tasks since they will not guarantee that the scheduler wait only on the I/O that is important first. The above example assumes only large files are written
Beckstrand Shelley M
International Business Machines - Corporation
Shin Christopher B.
LandOfFree
System and method for application influence of I/O service... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with System and method for application influence of I/O service..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and System and method for application influence of I/O service... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2471404