New Template.97

1
Adaptive File Caching in Distributed Systems
Ekow J. Otoo Frank Olken Arie Shoshani
2
Objectives

• Goals
• Develop a coordinated optimal file caching and
  replication of distributed datasets
• Develop a software module, called Policy Advisory
  Module (PAM) as part of Storage Resource Managers
  (SRMs) and other grid storage middleware
• Examples of application areas
• Particle Physics Data Grid (PPDG)
• Earth Science, Grid (ESG)
• Grid Physics Network (GriPhyN).

3
Managing File Requests at a Single Site
Multiple Clients Using a Shared Disk for
Accessing Remote MSS
Other Sites
Mass Storage System
Storage Resource Manager
Network
File requests
Queuing And Scheduling
Policy Advisory Module
Shared disk
4
Two Principal Components of Policy Advisory
Module

• A disk cache replacement policy
• Evaluates which files are to be replaced when
  space is needed
• Admission policy for file requests
• Determines which request is to be processed next
• e.g. may prefer to admit requests for files
  already in cache
• Work done so far concerns
• Disk cache replacement policies
• Development of SRM-PAM Interface
• Some models of file admission policies

5
New Results Since Last Meeting

• Implementation of the Greedy Dual Size (GDS),
  replacement policy
• New experimental runs with new workloads.
• 6 month log of access trace from Jlab
• Synthetic workload with file sizes from 500K to
  2.14G Bytes
• Implementation of SRM-PAM simulation in OMNeT
• Papers
• Disk cache replacement algorithm for storage
  resource Managers on the grid SC2002.
• Disk file caching Algorithms that account for for
  delays in space reservation, file transfers and
  file processing. To be submitted to Mass Storage
  Conference
• A discrete event simulation model of a storage
  resource manager (To be submitted to
  SIGMMETRICS2002) .

6
Some Known Results in Caching (1)

• Disk to Memory Caching
• Least Recently Used (LRU) keeps the last ref.
  time
• Least Frequently Used (LFU) keeps reference
  counts
• LRU-K keeps last reference times up to a max of
  K.
• Best known result (ONeil et al. 1993)
• Small K is sufficient (K2, 3)
• Gain 5-10 over LRU depending on reference
  pattern
• Significance of a 10 saving in time per
  reference
• Improved response time
• In the Grid and Wide Area Networks, this
  translates to
• Reduced network traffic
• Reduce load at the source
• Savings in time to access files

7
Some Known Results in Caching (2)

• File Caching in Tertiary Storage to Disk
• Modeling of Robotic Tapes Johnson 1995, Sarawagi
  1995,..
• Hazard Rate Optimal Olken 1983
• Object LRU Hahn et al. 1999
• Web Caching
• Self Adjusted LRU Aggarwal and Yu 1997
• Greedy Dual Young 1991
• Greedy Dual Size (GDS), Cao and Irani 1997

8
Difference Between Environments

• Caching in primary memory
• Fixed page size
• Cost (in time) is assumed constant
• Transfer time is negligible
• Latency is assumed fixed for disk
• Memory reference is instantaneous
• Caching in Grid environment
• Large files with variable sizes
• Cost of retrieval (in time) varies considerably
• From one time instant to another even for the
  same file
• Files may be available from different locations
  in a WAN
• Transfer time may be comparable to the latency in
  a WAN
• Duration of file reference is significant and
  cannot be ignored
• Main goal is to reduce network traffic and file
  access times

9
Our Theoretical Result on Caching Policiesin
Grid Environment

• Latency delays, transfer delays and file size
  impact caching policies in the Grid
• Cache replacement algorithms, such as LRU, LRU-K
  do not take these into account and therefore are
  inappropriate
• The replacement policy we advocate is based on a
  cost-beneficial function computed at time t0 as

t0 is the current time, ki(t0) is
the count of references for file i up to max of
K Ci(t0) is the cost in time of accessing the
file i, Si is size of file i. fi(t0)
is the total count of references to the file i
over its active time T. t-K is the
time of the kth backward reference.

• Eviction candidate is one with minimum gi(t0)

10
Implementations from the Theoretical Results

• Two new practical implementations developed
• - MIT-K Maximum average Inter-arrival Time, an
  improved
• LRU-K.
• - MIT-K dominates LRU-K
• - Does not take into account access costs and
  file size
• - The main ranking function is
• - LCB-K Least Cost Beneficial with K backward
  references
• - does take into account retrieval delay and
  file size

11
Some Details of the Implementation Algorithms

• Evaluation of replacement policies with no delay
  considerations involves
• - a reference stream r1, r2, r3, , rN
• - a specified cache size Z, and
• - two appropriate data structures
• One holds information of referenced files and
• A second holds information about the files in
  cache but also allows for fast selection of
  eviction candidate.

Cache content as a priority queue (PQ)
LRU
Binary Search Tree (BST)
12
Implementation When Delays are Considered
When delays are considered each reference ri in
the reference stream has 5 event times time of
arrival, time when file caching starts, time
when caching ends, time when processing begins
and time when processing ends and file is
released.
BST, Active lifetime T
Varies with different policies
A vector of pinned files in cache
PQ of unpinned files in cache
13
Performance Metrics

• Hit Ratio

• Byte Hit Ratio

• Retrieval Cost Per
• Reference

14
Implications of Metric Measures

• Hit Ratio
• Measure of the relative savings as a count of the
  number of files hit
• Byte Hit Ratio
• Measure of the relative savings as the time
  avoided in data transfers
• Retrieval Cost Per Reference
• Measure of the relative savings as the time
  avoided in data transfers and in retrieving data
  from their sources

15
Parameters of the Simulations

• Real workload from Jefferson Natl. Accelerator
  Facility (JLab)
• A six month trace log of file accesses to
  tertiary storage
• Log contains batched requests
• Replacement policy used in JLab is LRU
• Synthetic workload based on JLab
• 250,000 files with large sizes uniformly
  distributed between 500K to 2.147 GBytes
• Inter-arrival time is exponentially distributed
  with mean 90 sec
• Number of references generated is about 500,000
• Locality of reference
• partition the references into random size
  intervals
• follows the 80-20 rule within each interval(80
  of references are on 20 of the files)

16
Replacement Policies Compared

• RND Random
• LFU Least Frequently Used
• LRU Least Recently Used
• MIT-K Maximum Inter-Arrival Time based on last K
  references
• LCB-K Least Cost Beneficial based on last K
  references
• GDS Greedy Dual Size

• Active lifetime of a file T is set at 5-days
• All results accumulated with variance reduction
  technique.

17
Simulations Results for JNAF Workload Comparison
of Hit Ratios

• Higher values represent better performance

• MIT-K and LRU give the best performance
• LCB-K, GDS and RND are comparable
• LFU is the worst

18
Simulations Results for JNAF Workload Comparison
of Byte Hit Ratios

• Higher values represent better performance

• MIT-K and LRU give slightly best performance
• All policies except LFU are comparable
• LFU is the worst

19
Simulations Results for JNAF Workload Comparison
of Average Retrieval Time Per Reference

• Lower values represent better performance

• LCB-K and GDS give the best performance
• MIT-K, LRU and RND are comparable
• LFU shows the worst performance

20
Simulations Results for Synthetic Workload
Comparison of Average Retrieval Time Per
reference

• Lower values represent better performance

• LCB-K clearly gives the best performance although
  not significantly better than GDS
• LFU is still the worst
• Hit ratio and Byte Hit Ratio are not good
  measures of caching policy effectiveness on the
  Grid

21
Summary

• Developed a good replacement policy, LCB-K, for
  caching in storage resources management on the
  grid.
• Developed a realistic model for evaluating cache
  replacement policies taking into account delays
  at the data sources, transfers and processing.
• Applied the model for extensive simulation of
  different policies under synthetic and real
  workloads of access to mass storage system in
  JNAF
• We conclude that two worthwhile replacement
  policies for storage resource management on the
  GRID are LCB-K and GDS.
• The LCB-K gives about 10 savings in retrieval
  cost per reference compared to the widely used
  LRU.
• The cumulative effect can be significant in terms
  of reduced network traffic and reduced load at
  the source.