Electrical computers and digital processing systems: multicomput – Computer-to-computer data routing – Least weight routing
Reexamination Certificate
2000-03-31
2004-04-06
Maung, Zarni (Department: 2126)
Electrical computers and digital processing systems: multicomput
Computer-to-computer data routing
Least weight routing
C706S010000
Reexamination Certificate
active
06718358
ABSTRACT:
FIELD OF THE INVENTION
The present invention relates generally to the performance of networked systems and, more particularly, to automated techniques for improving the performance of networked systems.
BACKGROUND OF THE INVENTION
There has been a tremendous growth in the complexity of distributed and networked systems in the past few years. In large part, this can be attributed to the exploitation of client-server architectures and other paradigms of distributed computing.
Our interest is in automated techniques for improving the performance of heterogeneous distributed systems. Such systems are comprised of a variety of components including: computing elements, operating systems within the computing elements, middleware that links together computing elements, and applications that span computing elements. Herein, we use the term “target” to refer to the system, subsystem, or element that is being manipulated to improve its performance.
An initial question may be raised such as “why not solve performance problems by using more hardware, such as faster processors and more memory?” Sometimes, this is effective, at least up to the point where cost becomes a major issue. However, applying this approach in practice requires identifying resource bottlenecks, which requires some thought. Also, more hardware typically does not resolve logical bottlenecks, such as those due to locking or improper settings of task priorities.
The concept of “tuning” seeks to improve service levels by adjusting existing resource allocations. Doing so requires access to metrics and to the controls that determine resource allocations. In general, there are three classes of metrics: (1) “configuration metrics” that describe performance related features of the target that are not changed by adjusting tuning controls, such as line speeds, processor speeds, and memory sizes; (2) “workload metrics” that characterize the load on the target, such as arrival rates and service times; and (3) “service level metrics” that characterize the performance delivered, such as response times, queue lengths, and throughputs.
“Tuning controls” are parameters that adjust target resource allocations and hence change the target's performance characteristics. We give a few examples. Lotus Notes, an e-mail system and application framework, has a large set of controls. Among these are: NSF_BufferPoolSize for managing memory, Server_MaxSessions for controlling admission to the server, and Server_SessionTimeout for regulating the number of idle users. In Web-based applications that support differentiated services, there are tuning controls that determine routing fractions by service class and server type. MQ Series, a reliable transport mechanism in distributed systems, has controls for storage allocations and assigning priorities. Database products (e.g., IBM's DB/2) expose controls for sort indices and allocating buffer pool sizes.
To determine the effect of tuning adjustments, there must be a model that relates workload levels and the settings of tuning controls to the service levels that will be achieved. We refer to this as the target's system model, or just “system model.” It is difficult to acquire and maintain a system model, especially since application software can change dramatically from release to release (even between patch levels). Thus, system models are usually informal and imprecise.
In the existing art, tuning typically involves the following steps: (1) collect data; (2) use the system model to determine how tuning controls should be adjusted; and (3) goto step (1).
There are many challenges here. First, as noted previously, acquiring and maintaining the system model is difficult. Second, the controls are complex and often impact service levels in nonlinear ways. This makes it challenging to select the tuning controls to adjust as well as to determine what the settings of these controls should be. Third, the above scenario only considers current workloads. It may well be that by the time the tuning controls are adjusted, the workload will have changed so that new adjustments are necessary.
Because the expertise required for manual tuning is scarce, many have pursued an automated approach. A variety of target-specific or “customized automated tuning systems” (CATS) have been developed. Examples include systems by: (1) Abdelzaher and Shin, as described in “End-host Architecture for QoS-Adaptive Communication,” IEEE Real-Time Technology and Applications Symposium, Denver, Colo., June 1998, the disclosure of which is incorporated by reference herein, who control quality of service for the delivery of multimedia using task priorities in the communications subsystem; and (2) Aman et al., as described in “Adaptive algorithms for managing a distributed data processing workload,” IBM Systems Journal, Vol. 36, No 2, 1997, the disclosure of which is incorporated by reference herein, who provide a means by which administrators specify response time and throughput goals to achieve in MVS (Multiple Virtual Storage) systems using MVS-specific mechanisms to achieve these goals.
CATS require that metrics and tuning controls be identified in advance so that mechanisms for their interpretation and adjustment can be incorporated into the automated tuning system. Thus, CATS construction and maintenance still require considerable expertise. With the advent of the Internet, software systems and their components evolve rapidly, as do the workloads that they process. Thus, it may well be that automated tuning systems must be updated on a rate approaching that at which tuning occurs. Under such circumstances, the value of automated tuning is severely diminished.
Since customized automated tuning systems are difficult to build and maintain, it would be highly desirable to have a generic automated tuning system. Thus, instead of requiring experts to incorporate detailed knowledge of the target, such a generic automated tuning system may learn the target's performance characteristics. This may include having such a generic automated tuning system exploit prior knowledge of the target system, when such knowledge is available, reliable, and durable. As will be explained herein, the present invention provides such a generic automated tuning system and methodology.
As will be further explained in accordance with the present invention, a starting point in building such a generic automated tuning system is to construct a generic system model. Prior art in learning system models has largely focused on neural network approaches such as those in U.S. Pat. Nos. 5,893,905 to Main et al.; 5,461,699 to Arbabi and Fischthal; and 5,444,820 to Tzes and Tsotras, the disclosures of which are incorporated by reference herein. More related to objectives of the present invention is the work in U.S. Pat. No. 5,745,652 to Bigus, the disclosure of which is incorporated by reference herein, that describes a target-independent approach to automated tuning. This is accomplished by having a neural network that is trained off-line to learn the system model. In on-line operation, the system model is used in combination with a second neural network, the controller, to learn control actions.
The foregoing address two issues associated with constructing a generic automated tuning system, as is provided in accordance with the present invention: the generic system model and tuning control estimation. However, as will be explained, many other considerations are necessary as well.
First, existing art for target-independent automated tuning does not consider architectural support for access to the metrics and controls. Realizing generic, automated tuning requires well defined interfaces so that a generic automated tuning system can access the data required from the target. Previous work has ignored these considerations.
Second, the search for appropriate settings of tuning controls is facilitated by exposing information about the semantics of metrics and the operation of tuning controls. In particular, it is helpful for the target to place metrics into t
Bigus Joseph Phillip
Hellerstein Joseph L.
Squillante Mark S.
Thathachar Jayram S.
International Business Machines - Corporation
Maung Zarni
Opie G. L.
Ryan & Mason & Lewis, LLP
Zarick Gail H.
LandOfFree
System and method for generic automated tuning for... 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 generic automated tuning for..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and System and method for generic automated tuning for... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3263755