Electrical computers and digital data processing systems: input/ – Input/output data processing – Peripheral monitoring
Reexamination Certificate
2001-01-16
2004-02-03
Elamin, Abdelmoniem I (Department: 2182)
Electrical computers and digital data processing systems: input/
Input/output data processing
Peripheral monitoring
Reexamination Certificate
active
06687765
ABSTRACT:
FIELD OF THE INVENTION
This invention pertains generally to input/output device performance tuning, and more particularly to a RAID device controller that is explicitly tuned to provide high-performance without imposing significant overhead during normal input/output operations.
BACKGROUND
Modern computers, particularly computers operating in a server environment, benefit from a large, fault-tolerant data storage system. Hard drives in all computer systems are susceptible to failures caused by temperature variations, head crashes, motor failure, controller failure, and changing voltage conditions, among other causes. To improve reliability and protect the data in data storage systems, many data storage systems use a redundant array of independent disks (RAID) operated by a disk array controller. In a RAID storage system, data storage redundancy is achieved by mirroring a copy of the data on one disk drive onto another disk drive (e.g. RAID
1
), or by storing the data on a plurality of disk drives along with parity data, that permits the original data on any single drive that may fail to be reconstructed on the basis of the data and parity stored on the remaining disk drives (e.g. RAID
3
, RAID
4
, RAID
5
, as well as other RAID and non-RAID drive systems). RAID storage systems and RAID levels are well known in the art and not described in greater detail here.
Conventional RAID systems typically consist of several individual disk controllers combined with a rack or other physical configuration of drives to provide a fault-tolerant data storage system that is directly attached to a host computer. The host computer is then connected to a network of client computers to provide a large, fault-tolerant pool of storage accessible to all network clients. Typically, the disk array controller provides the brains of the data storage system, servicing all host requests, storing data to multiple drives, such as, for example, RAID drives, caching data for fast access, and handling any drive failures without interrupting host requests.
Caching is a data storage mechanism which is based on the assumption that a data fetch operation to a storage device, such as to a disk drive or RAID storage subsystem, to retrieve one item of data may be followed by a subsequent fetch operation to the same storage device to retrieve a related item of data which may frequently be a data item stored on an adjacent or sequential portion of the storage media of the storage device. Therefore, during the first read operation, not only is the required data item read, but an amount of additional data that is associated with the required data is also read. Typically, the amount of additional data read is determined by the cache memory size, the bandwidth of the communication channel between the storage device (and/or the controller) and the cache, some predetermined amount that is arbitrarily set, or the like factors as are known in the art. For example, when reading a portion of a document for a word processing application, more of the document is retrieved from the hard disk and stored in the cache than is actually required for display to the user at that time, in anticipation that the user may scroll the document and required additional data to be retrieved and displayed. The retrieval of data from the cache, usually a fast electronic Random Access Memory (RAM), is much quicker than retrieval from disk storage. Similar efficiencies may typically be realized when reading program instructions for execution, as many (though not all) instructions will execute or at least be launched in order. Additionally, branches to an instruction, even if remote from the instruction, are often cached in case the branch becomes valid. Therefore, data or other information caching (such as instruction caching) may frequently, though not necessarily always, improve input/output performance. Numerous caching schemes are known in the art and not described in greater detail here.
An input/output profile (I/O profile) of an application or workload is a characterization of the application that typically includes the type and frequency or number of I/O operations for that application. For example, the application may issue several different types of I/O operations including: small random I/O requests, small block size sequential I/O requests, large block size non-sequential I/O requests, and large block size sequential I/O requests. These types are merely examples and not intended to be limiting, as it will be appreciated by those workers having ordinary skill in the art that there may generally be a continuum of block sizes between “small” and “large” and that a “small” block for some data bases or applications may be a “large” block for others.
Typically, the performance accomplished by an I/O controller, such as a RAID controller, and its controlled devices may be heavily dependent on the input/output (I/O) profile of its workload and the configuration or tuning of the controller, particularly relative to the use or non-use of caching. We here consider three exemplary operational I/O situations having different I/O profiles for purposes of illustration. In a workload having predominantly small-block random I/O requests, no particular advantage is gained by employing caching algorithms as random I/O requests will not typically exhibit any hits into the cache. In a workload where random I/O requests predominate, the performance is dependent on the overhead in the controller processing the I/O requests and the speed and number of disk drives attached to it. From the controller standpoint, the overhead is typically a function of the speed of instruction execution of the processor and the efficiency of its algorithms.
By comparison, small-block sequential I/O operations (for example, block sizes between about 512 and about 8K to 16K bytes) may greatly benefit from read-ahead caching because small-block sequential read operations with read-ahead caching tends to minimize the number of requests to the disks or other storage device. Increasing the number of requests to the disks tends to degrade I/O performance, particularly where the number of requests becomes large. In such instances, the use of the read-ahead caching algorithms will result in a significant performance gain over non-caching algorithms as the number of I/O requests to the disks are minimized. However, even for sequential read operations, a non-caching procedure may typically provide better performance than a caching procedure when the block size is large, as in these circumstances the performance is determined primarily by the bandwidth of the internal data paths which limits the overall rate at which data can be moved, rather than being primarily determined by caching.
Therefore, accurately determining the input/output profile of an application or workload, as well as the operational profile of the devices and/or the device controller, and configuring the device and/or device controller to optimize the types of I/O operations can greatly enhance system performance. The benefits may be particularly dramatic when a RAID controller is optimally configured for the I/O profile of the workload or application.
The performance of a RAID controller, usually specified in terms of input/output operations per second (IO/s), bandwidth (MB/s), or I/O completion time, is commercially important to the acceptance and sales of the controller in the marketplace. Competitive performance in state-of-the art RAID controllers has heretofore been achieved through several means. These means include setting various RAID controller and/or controlled device parameters (such as for example, setting or adjusting the RAID stripe size), whether read-ahead caching is enabled or disabled, whether direct I/O or cached I/O is used, whether write-back (WB) or write-through (WT) caching is used, and setting cache aging, cache size and the like.
Measured parameters may, for example, include I/O locality and I/O time. Tuned parameters may include, for example: (i) cached I/O versus direct I/O, and cache size, eac
Cassidy Bruce M.
Surugucchi Krishnakumar Rao
Ananian R. Michael
Dorsey & Whitney LLP
Elamin Abdelmoniem I
International Business Machines - Corporation
LandOfFree
System, method, and computer program for explicitly tunable... 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, method, and computer program for explicitly tunable..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and System, method, and computer program for explicitly tunable... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3337920