Mobile and Internet Systems Group A Semantic-based Cache Replacement Algorithm for Mobile File Access Sharun Santhosh and Weisong Shi Department of Computer

Mobile and Internet Systems Group

A Semantic-based Cache Replacement Algorithm for Mobile File Access

Sharun Santhosh and Weisong Shi

Department of Computer ScienceWayne State [email protected]

http://mist.cs.wayne.edu


Motivation

The Future Staying connected anywhere, anytime will become a

reality

How ? Cable modem or DSL connection at home High speed Ethernet network at work or school Satellite network in the car WiFi network at the airport or the neighborhood coffee

shop

Challenges Effectiveness - Adapt to the various underlying

connectivity Convenience - Adaptation should be transparent to the

user Security – secure access in resource constraint devices


Heterogeneous Environment

Wide Area Wide Area Network (WAN)Network (WAN)

9.6 Kbit/s <2Mbs9.6 Kbit/s <2Mbs• Voice• SMS• e-Mail• Web browsing

• mCommerce• Internet access• Document transfer• Low/high quality video

GPS

Local Area Network wLAN

802.11a,b,g802.11a,b,g

LAN

<100Mbs<100Mbs• Access•“hot spots”•LAN equivalent

WirelessBridge

WorkgroupSwitches

Personal Area Network (PAN)

<1Mbs<1Mbs• Access•Synchronization•10 Meters

Bluetooth

GSM/CDMA


Adaptive Comm

unication

Optimization (Fractal)

Conn

ectio

n Vi

ew b

ased

Secu

re a

nd T

rans

pare

nt

Reco

nnec

tion

Semantic based Caching

Our Solution

CEGORClosE and Go,

Open and Resume


Roadmap

Motivation

Caching

Semantic-based Caching

Simulation Results

Conclusion


Caching

Three basic steps involved in accessing data anywhere and anytime. Retrieve the files from the server Work on them locally Write the changes back to the server

A Cache optimizes this process Reduce frequency of disk operations performed Reduce frequency of requests to the fileservers Reducing network load

Problem being addressed Minimization of CommunicationMinimization of Communication


Why Study Caching?

It has been studied extensively yet LRU is the most commonly used algorithm Used in NFS, AFS, Sprite, CODA and most operating

systems buffer caches

Why ? It’s simple to implement. Cache misses are acceptable in existing systems. Number of files replaced do not matter high hit ratio vs. # of replacement

But in a heterogeneous environment Each miss implies additional communication Storage of work (when in a weakly connected or

disconnected state) Cannot assume a reliable link exists with the server


Usage Scenario

“Imagine a field engineer is accessing layout diagrams for a faulty electricity sub-station, half way through communications go down. A cache MISS may cause several minutes delay, perhaps longer, e.g.,

Which was the 10,000 volt cable?”


Is simple caching (LRU) enough??


Goals

Caches for distributed file systems, that operate across heterogeneous networks must

Provide the hit rates of conventional caches that operate over homogenous networks

Minimize communication overhead, i.e., minimize replacements which mean increased file availability and


Our Approach to Caching

File access patterns aren’t random

A semantic relationship exists between two files in a file access sequence User behavior Program execution

We define and investigate two kinds of such relations Inter-file relations Intra-file relations

We introduce the notion of eviction index for each cached item


Outline

Motivation

Caching


Simulation Results

Conclusion


Inter-file relations

Analysis of DFS traces


Inter-file relations

An inter-file relationship exits between two files i and j, if i is the next file opened following j being closed. File j is called file i’s precursor.

Xi - represents the number of times file i is accessed.

Ti - represents the time since the last access to file i.

Yj - represents the number of times file j precedes file i.


Intra-file relations

An intra-file relationship is said to exist between two files i and j if they are both open before they are closed.

Intra-file relations are based on shared time Si,j

defined below Where O(i) and C(i) are the time at which file i was

opened or closed respectively

i

j

C

C

O

O

Sij


Intra-file relations

Ti - represents the time since the last access to file i.

Tj - represents the time since the last access to file j where j is open before i is closed.

Si,j - represents the shared time of file i with respect to file j where i is closed before j.

Stotal - represents the total shared time with all files that are open before i is closed


Inter + Intra


Workload

DFS Traces from CMU were utilized during the simulation


Implementation

Seven replacement algorithms RR – Round Robin LRU – Least recently used LFU – Least frequently used GDS – Greedy dual size INTER – based only on inter-file relations INTRA – based on intra-file relations Both – based on both intra and inter file relations

Varying cache sizes 10KB, 25KB, 50KB, 100KB, 500KB …

Seven traces Simulator maintains a cache (hashlist), open list

(list of currently open files), close list (list of files that are closed).


Structure of simulator


Simulator pseudocode


Outline

Motivation

Caching


Simulation Results

Conclusion


Hit rates of all algorithms

INTER/BOTH


INTRA

Replace attempts of all algorithms

INTER/BOTH


Performance – DFS Traces


Performance – DFS Traces


Performance


The Need For File System Tracing

Traces haven’t been collected periodically enough to reflect present day usage activity

Publicly available traces such as traces collected at the disk driver level or web proxy traces do not give us relevant information on file system workload.


New Open

Data Logger

Trace Module

System Call Interception

Standard Library

int fopen(char *name,char *mode)

USER SPACE KERNEL SPACE

Original

Open

Original

Close

Open Close

System Call Table


Analysis Summary

Most files were opened for less than a hundredth of a second

Majority of files are accessed only a few times. There is a small percentage of very popular files

Majority of files are less than 100KB in size. Large file can be very large (heavy tail)

Almost half the accesses repeat within a short period of initially occurring

File throughput has greatly increased due to presence of large files

Majority of files accessed have a unique predecessor


MIST traces – Hit Rates


MIST Traces – Files Replaced


MIST Traces –Byte Hit Rate


Summary

We have presented a semantic-based caching algorithm and shown that it performs better than conventional caching approaches in terms of hit ratio and byte hit ratio

We have also shown that it does this performing far

fewer replacements

Compared to prevalent replacement strategies that ignore file relations and communication overhead, this approach would seem to better suit distributed file systems that operate across heterogeneous environments


Future Work

Collecting more state-of-the-art distributed file systems traces

Applying the cache replacement algorithm into a real wireless file system in computer-assisted surgery application

Investigating the idea into more general applications, such as mobile database access, etc.


Questions & Comments?

[email protected]

http://mist.cs.wayne.edu

Documents

Mobile and Internet Systems Group A Semantic-based Cache Replacement Algorithm for Mobile File Access Sharun Santhosh and Weisong Shi Department of Computer