280 likes | 440 Views
Performance Prediction Using Program Similarity. Aashish Phansalkar & Lizy K. John. The University of Texas at Austin. Outline. Motivation and Objectives Methodology Experimental results Conclusion Future work. Motivation (1): Simulation is costly.
E N D
Performance Prediction Using Program Similarity Aashish Phansalkar & Lizy K. John The University of Texas at Austin
Outline • Motivation and Objectives • Methodology • Experimental results • Conclusion • Future work
Motivation (1): Simulation is costly • A computer architect or a designer has to simulate multiple customer applications • Simulations take very long due to the complexity of modern microprocessor designs
Motivation(2): Making a decision based on benchmark scores • Customers often use benchmarks to make a decision about buying computer systems • The application program they use often, may not be a part of the benchmark suite • Customers can use benchmarks as representatives of their application programs • Predict performance of their application based on the already available performance data of benchmarks
Objective • A quantitative method to estimate performance without running cycle accurate simulation • Use the knowledge of similarity between a customer’s application program and known benchmark programs to develop a quantitative approach to predict performance
Outline • Motivation and Objectives • Methodology • Experimental results • Conclusion • Future work
Overview Known cases New Case Repository of Benchmarks Customer application Measure similarity Predicted performance
Program characterization • Instruction mix • Percentage of different types of instructions e.g. percentage of memory references, percentage of branch instructions • Control Flow • % Taken branches • % Forward branches • % Forward taken branches • Basic Block Size (Number of instructions between two branches) • Register Dependency Distance • Data and instruction temporal locality of program • Data and instruction spatial locality of program
Register dependency distance ADD R1, R3,R4 MUL R5,R3,R2 ADD R5,R3,R6 LD R4, (R8) SUB R8,R2,R1 Read After Write Dependency Distance = 4 Measure Distribution of % of dependency distances for following set of ranges. 1, 2, 3-4, 5-8, 8-16, 16-32, greater than 32 Normalized count for each range of dependency distance forms a metric
Data and instruction temporal locality Memory reuse distance 2004, 2008, 4008, 2000, 1080,2004,4008 Reuse Distance = 4 Reuse Distance = 3 • Computing reuse distance for a trace of byte addresses is very computation and space intensive • Reuse distance for a block of 16, 64, 256, 4096 bytes • Temporal locality metrics (tlocality) = Wt. average reuse distance
Data and instruction spatial locality • Spatial locality metrics are derived from the temporal locality metrics • As the block size increases, programs with good spatial locality will show lower values for tlocality for higher block sizes • Spatial Locality = tlocality64 / tlocality16 tlocality256 / tlocality16 tlocality4096 / tlocality16
Methodology Overview Microarchitecture independent metrics for known benchmarks Microarchitecture independent metrics for the customer application Measure program similarity Similarity information Prediction of target metric for new application (2 methods) Predicted value of target metric
Measuring Similarity (1) • Distance between two programs in the workload space is the measure of their similarity • We assume that similarity between two programs is inversely proportional to the Euclidean distance between them
Measuring similarity (2) • The workload space is made of many workload characteristics and so its dimensionality is very high • Inherent characteristics are highly correlated • Euclidean distance measured using these characteristics will be biased • The correlated variables will add twice to the distance as the independent variables • Use Principal Components Analysis (PCA)
Method 1: Predicting performance using weights • Compute distance of similarity from program X to each benchmark program dx1, dx2, dx3…dxn in the PC space • Calculate weights w1, w2, …. w1 User program X w2 benchmarks
Method 2: Predicting performance using clustering • Measure all the inherent characteristics for the benchmarks and user program X • Cluster all the programs based on the inherent characteristics and find optimal clusters User program X benchmarks
Outline • Motivation and Objectives • Methodology • Experimental results • Conclusion • Future work
Experiments • Used integer programs from SPEC CPU2000 suite to demonstrate the use of Method 1 and Method 2 described • Prediction of speedup • Used all the workload characteristics to form the workload space • Prediction of cache miss-rates • Used only the data locality characteristics to form the workload space
Predicting speedup(1) • Experiment: Predict performance (speedup) of bzip2 using benchmarks from SPEC CPU2000 suite • Assume that bzip2 is the customer application • Performance of SPEC CPU2000 benchmarks is known Speedup for each benchmark program on a machine (from the scores reported on the SPEC website)
Predicting speedup(2) Method 1:Predicting speedup using weights Machine name: SGI-Altix 3000 (1500MHz, Itanium 2)
Predicting speedup (3) Method 1: Predicting speedup using weights For 50 different machines the error in predicted speedup
Predicting speedup (4) Method 2: Predicting speedup using clustering The average error in predicting the speedup over all machines for bzip2 is 20.29%
Prediction of data cache miss rates(1) Method 1: Using weights for prediction Note: Assume every program to be a customer application one at a time
Prediction of data cache miss rates(2) Method 2: Using clustering for prediction Note: Assume every program to be a customer application one at a time
Outline • Motivation and Objectives • Methodology • Experimental results • Conclusion • Future work
Conclusion • Demonstrated two simple methods to predict performance • Used SPEC CPU2000 as an example to predict performance. • The accuracy of prediction depends on two factors: • How well the workload characteristics correlate to performance • Is there a program similar to the customer application in the repository of known programs
Future Work Two main items on the TO DO list: • To add more programs to the repository and validate the results • To calibrate the measure of similarity (distance) in workload space to the error in the target metric space.