Application-level Prefetching

CS656

semester projectPeixan Li, Jinze Liu, Hexin Wang

Outline

MotivationSolutionImplementationEvaluation Test PlanConclusion

Motivation

Get Better Performance on Remote Data Access over DOS Random Data Access Sequential Data Access

e.g., Video-on-Demand Data Transfer

Problems Low Bandwidth Network

Wait for your data whenever you need it most

General-Purpose OS Inappropriate Scheduler (RoundRobin) not addressing timing constraints

Solution(I) -- Client Side

Application level Prefetching Cache Although the fact -- Low Bandwidth Network Prefetching Cache can reduce data access time

General Data ServiceHistory-based prefetching

Video-on-DemandSequential Prefetching

Solution(II) -- Server Side

Admission Control Avoiding Server Overload All the admitted tasks can be satisfied using current

scheduling algorithm

Server Level Scheduling algorithm Isochronous Tasks can run first before GP Tasks

Implementation

Client/Server ModelClient Side

Prefetch via Data Compression Cache Management

Server Side Admission Control Scheduling Algorithm

Client/Server Model

V/A PlayerNormal Application Admission Control

Schedule Control

Service Thread

Service Thread….

File Server

Controller

Prefetch Thread

Cache

Controller

Prefetch Thread

Cache

Client 1 Client 2 Server

Peixian...

Prefetch via Data Compression

Based on data compression techniquesWhy is D/C useful for prefetch

Basic law: To represent more common events with short codes and less common events with longer codes

Must be good at recording history and predicting future data

Be particularly good for databases and hypertext systems

History-based Prefetch

We use Ziv-Lempel algorithm Simple but very good Predict based on a probabilistic

history tree e.g. “aaaababaabbbab”

=> (a)(aa)(ab)(aba)(abb)(b) Sequential prefetch is used when

lack of history Prefetch thread is activated once a request is finished History tree need to be rebuilt before it becomes too large

Sequential Prefetch

Two kinds of interfaces provided client module Hread() is for history-based prefetch Sread() is for sequential prefetch More kinds of reads can be added, e.g. real-time

When sequential prefetch is used No history is needed Only future data need to be cached Semaphores are used to synchronized cache-read and

cache-write

Cache Management

Cache size dynamically grows and shrinks With default size and maximum limit In order to use memory efficiently

In order to provide better performance Use LRU replacement algorithm

Simple but good enough

No consistency issue since we only have read-only access

Admission Control(I)

Basic Assumption Isochronous Tasks

Real-time periodic tasks• MPEG-1 requires about 1.5 MbitsPS

• MPEG-2 or MPEG-4 requires about 5-10MbitsPS

Require performance guarantee for throughput, bounded latency.

General-Purpose TasksPreemptible tasksSuitable for low-priority background processing

Admission Control(II)

If a new isochronous task is to be admitted All the previous tasks must be satisfied whenever new task

is taken into account or not The new task can be satisfied under current workload

High frequency tasks run before low frequency tasksA periodic task can be satisfied means it can finish

within each period. I.e., Real Execution Time <= Period

Admission Control(III)

To admit a new Isochronous Task I.e.

n -- Total number of isochronous tasksCi -- An execution time per period of task iTi -- Period of isochronous task i

Disadvantage General-Purpose Tasks may suffer from starvation

Admission Control(V)

E.g. Task1(C1 = 6, T1 = 10); Task2(C2 = 6, T2 = 20);

T1

T2

T3

0 10 20 30 t

Scheduling algorithm

Schedule Algorithm Isochronous requests scheduled using rate monotonic

The higher frequency, the higher priority

Normal file requests scheduled with round robinCan be preempted by isochronous tasks

Jinze...

Test Plan

Test programs with different access patterns Sequential remote multimedia access. Simulated tree-like web document access. Simulated database access. Random remote file access.

To test prefetching performance with different test programs.

To test server performance with concurrent requests of different applications

Evaluation

Performance comparison -- Yes/No prefetching Cache Hit Rate Received throughput

Server performance with different tasks Correctness of Admission Control. Measurement of capacity

Unsolved Problems

Cache cannot be shared between different applications.

Cache data is lost after the termination of application program.

Cache is read-only.

Conclusion

We’ve simulated a client/server model to support application-oriented isochronous prefetching.