Download Performance Engineering: Ensuring Program Performance and Addressing Bottlenecks and more Study notes Software Engineering in PDF only on Docsity!
Evidence: the midterm. My initial policies don't properly support the outcomes I desire. Evidence: people wishing to use the course to apply to the MS program. The course has been used for unforeseen purposes. Evidence: too much grading to do, no help available. In practice, the structure I proposed isn't sustainable. As of now, this course as an experiment is "at risk": Problem: assignments too abstract. Solution: make assignments concrete. Need: skill building Problem: group assignments do not serve. Solution: individual assignments. Need: certification of individual skills Problem: essays are difficult to grade. Solution: exercises in practice are easier Need: sustainability Therefore, for the survival of the course as a concept , it is necessary to rethink the course to address these needs: Some straight talk on the future of the course Monday, November 05, 2012 8:21 AM Docsity.com
Solution: exercises in practice are easier to complete and to grade. Docsity.com
A Service Level Objective (SLO): defines what is "fast enough". A Service Level Agreement (SLA): defines economic penalties for not meeting SLOs. Some important language from Service- Oriented Architecture (SOA) Some important language from Service Oriented Architecture Monday, November 05, 2012 9:02 AM Docsity.com
"Web service requests should respond in less than 1 second of real time. An appropriate SLO: "If less than 90% of requests meet the SLO in a month, the customer doesn't pay anything for the service." A related SLA: SLOs and SLAs Monday, November 05, 2012 8:58 AM Docsity.com
Performance isn't free. But neither is lack of performance. Often, performance engineering is a matter of balancing the cost of performance against its value. The point of SLOs and SLAs The point of SLOs and SLAs Monday, November 05, 2012 9:03 AM Docsity.com
Factoring: what factors contribute to performance Performance measurement: how we document how programs perform Performance analysis: how we determine the performance problems for software. Performance patterns: how we address common performance problems. Performance engineering Performance engineering Monday, November 05, 2012 8:44 AM Docsity.com
Performance problems are usually expressed in terms of bottlenecks. A bottleneck is a specific performance problem that -- if addressed -- most effectively speeds up total execution. You might also hear a bottleneck defined through specifying the "critical resource". The concept of a bottleneck The concept of a bottleneck Monday, November 05, 2012 9:09 AM Docsity.com
CPU-bound: the critical resource is the speed of the processing unit. Cache-bound: the critical resource is the size of the CPU cache. Memory-bound: the critical resource is the speed of the (front-side) memory bus. Disk-bound: the critical resource is the speed of (local) disk access. Network-bound: the critical resource is the speed of the connected network/internet. Service-bound: the critical resource is the speed of response of a remote service. Kinds of bottlenecks: Kinds of bottlenecks Monday, November 05, 2012 9:11 AM Docsity.com
times command: tells how long a program ran (in cpu time and real time). getrusage system call: indicates time used, memory used. profiling: gives detailed resource consumption of functions. Performance measurement: Performance measurement Monday, November 05, 2012 8:46 AM Docsity.com
Compile a special statistics library into the code. Keep statistics on CPU consumption in each subroutine. Profiling gcc - pg file.c ./a.out # store results gprof a.out # report results Profiling Monday, November 05, 2012 10:35 AM Docsity.com
for (i=0; i<10; i++) foo(i); Change: foo(0); foo(1); foo(2); foo(3); foo(4); foo(5); foo(6); foo(7); foo(8); foo(0); To: A common form of performance optimization: loop unrolling: The actual mechanics of the loop slow down execution (i<10, i++) Why? Case study: loop unrolling Monday, November 05, 2012 11:02 AM Docsity.com
In truly performance-critical code, it is often advantageous to eliminate if statements and replace them with indirections. Eliminating "if" statements if (something(x)) foo(x); Example: Becomes foo(redirect(x)) Where x if something(x); a random other place if! something(x); redirect(x) has two values: The trick is to do this without invoking the "if" statement at all, e.g., by making redirect(x) involve only arithmetic. For example, in a profiler, it is cheaper to measure everything (and then throw some statistics away) than to measure selectively, because the "if" statement slows down the code. Advanced performance optimization: eliminating if statements Monday, November 05, 2012 11:04 AM Docsity.com
If runtime is not solely dependent upon program design.
○ If external factors contribute. Often gprof fails to provide anything useful The complex cases The complex cases Monday, November 05, 2012 10:56 AM Docsity.com
- Problem: clock resolution is 1/100 second. Solution: scaffold test, repeat test 1000 times, divide by 1000. Problem: performance measurement of inaccessible subsystems (disk, network, ...).
Resource latency: time spent waiting for resources. Concurrency latency: time spent waiting for concurrent (unrelated) tasks to finish. Solution: Wallclock time - Execution time = Resource latency + Concurrency latency Typical approach: isolate test on quiescent system, then concurrency latency is 0. Performance measurement problems and solutions: Performance measurement problems and solutions Monday, November 05, 2012 9:29 AM Docsity.com
