Generating Streaming Access Workload for Performance Evaluation Computer

Generating Streaming Access Workload for Performance Evaluation Computer Science Shudong Jin 3 nd Year Ph. D. Student (Advisor: Azer Bestavros)

Project Overview Computer Science q This project aims to develop a Generator of Internet Streaming Media Object access workloads (GISMO) q Why develop GISMO? =Streaming access of emerging Internet streaming application (e. g. , video/audio on Web) has unique characteristics: - High bandwidth requirement Long duration (seconds to hours) Variable bit-rate (VBR) burstiness Timeliness and user-perceived quality are important =There is no streaming access workload generator - Workload generation is important for performance evaluation of Internet streaming content delivery techniques

GISMO: Characteristics Computer Science

GISMO: Modeling Computer Science q Modeling Request Arrival Process =Popularity distribution - Zipf-like distribution models the skewed request frequency of the streaming media objects. P ~ r- , 0< <1, where P is the access frequency, r is the rank of an object. =Temporal Correlation of Requests - Requests to the objects tend to arrive non-randomly. Pareto distribution models the correlated inter-arrival time. =Seasonal Patterns - Aggregated request arrival rate can exhibit seasonal patterns (hourly, daily, weekly etc). GISMO users can define such diurnal patterns.

GISMO: Modeling Computer Science q Modeling Individual Requests =Object Size Distribution - Streaming media objects have a wide range of length. We use a power law to model it. =Partial Access Patterns - User interactions involves in streaming access. We use Pareto distribution to model the stop time. =Variable Bit-Rate - The bit-rate of streaming media objects has high variability. We use Pareto distribution to model the tail of VBR marginal distribution, and Lognormal distribution for the body.

GISMO: Modeling Computer Science q VBR self-similarity = The bit-rate of streaming media objects (e. g. , audio/video) exhibits long-range dependence. = The auto-correlation function decay slowly = Burstiness persists for long period, and implies the ineffectiveness of buffering q Generating self-similar process FGN = We use a random middle-point displacement algorithm q Transforming VBR marginal distribution = Gaussian hybrid Lognormal/Pareto distribution

GISMO: Functionality Computer Science q GISMO generates =A set of bogus streaming media objects, installed in the servers which mimic real servers =Requests to these objects, initiated by the clients which mimic real users q GISMO can be used for many purposes =Evaluating the performance of streaming media servers, e. g. , scheduling and I/O =Evaluating network protocols for streaming data transmission =Evaluating streaming data replication techniques (caching, pre-fetching, multicast merging, etc)

GISMO: Architecture Computer Science WWW Browser Requests TCP Media Player WWW Browser Requests Streaming Server Objects Network RTSP Media Player Requests UDP WWW Browser Web Server

GISMO: Use Case Computer Science q We have conducted a case performance study =Using GISMO to generate workloads =Evaluating proxy caching and server stream merging techniques =Showing that how the workload characteristics impact their effectiveness

GISMO: Use Case Computer Science How does popularity impact the effectiveness of proxy caching (left) and server merging (right)

Future Directions Computer Science =More client interactions in request streams, e. g. , VCR functionality =More correlations in streaming media objects, e. g. , Group-of-Picture Go. P correlation =Using GISMO in evaluating streaming content delivery techniques =Using GISMO in evaluating network protocols for streaming data transmission

Related Publications Computer Science q Shudong Jin and Azer Bestavros. Generating Streaming Access Workloads for Performance Evaluation and A Case Study. BU CS Technical Report, April 2001. q Shudong Jin and Azer Bestavros. Temporal Locality in Web Request Streams: Sources, Characteristics, and Caching Implications. Short paper appeared in ACM SIGMETRICS’ 2000; full paper appeared in MASCOTS’ 2000. q Paul Barford and Mark Crovella. Generating Representative Web Workloads for Network and Server Performance Evaluation. ACM SIGMETRICS’ 1998.