Quantifying Temporal and Spatial Quantifying Temporal and Spatial - - PowerPoint PPT Presentation

Florida State University, {fox, lojpur, Florida State University, {fox, lojpur, awang@cs.fsu.edu} awang@cs.fsu.edu} 1 1

Quantifying Temporal and Spatial Quantifying Temporal and Spatial Localities Localities

Florida State University fox, lojpur, awang@cs.fsu.edu

Florida State University, {fox, lojpur, awang@cs.fsu.edu} 2

Quantifying Localitites Quantifying Localitites

• Describe workloads

Describe workloads

_ _ Compare different workloads

Compare different workloads

_ _ Create accurate synthetic workloads

Create accurate synthetic workloads

• Understand locality transformations

Understand locality transformations

_ _ How are workloads transformed by caching?

How are workloads transformed by caching?

_ _ Do workloads properly stress data path components?

Do workloads properly stress data path components?

_ _ What opportunities remain for further data path optimizations?

What opportunities remain for further data path optimizations?

Florida State University, {fox, lojpur, awang@cs.fsu.edu} 3

Current Metrics Current Metrics

_ _ Cache hit ratios

Cache hit ratios

• Measures the effectiveness of caching

Measures the effectiveness of caching

• Does not cleanly separate temporal and spatial locality

Does not cleanly separate temporal and spatial locality

• Does not show transformations

Does not show transformations

• Only applies to cache

Only applies to cache

_ _ Reference distance

Reference distance

• Reference distance is the number of unique blocks referenced

Reference distance is the number of unique blocks referenced before accessing the same block before accessing the same block

_ _ Block distance

Block distance

• Block Distance is the difference between block locations on disk

Block Distance is the difference between block locations on disk

Florida State University, {fox, lojpur, awang@cs.fsu.edu} 4

New Metrics New Metrics

_ _ Affinity

Affinity

• Builds on block and reference distance

Builds on block and reference distance

• 0%-100% easy to understand

0%-100% easy to understand

• Less sensitive to generations of hardware

Less sensitive to generations of hardware

Florida State University, {fox, lojpur, awang@cs.fsu.edu} 5

Example use Example use

Web-trace Synthetic

Shuffled references, same inter-arrival rate, reference distribution

Florida State University, {fox, lojpur, awang@cs.fsu.edu} 6

Future Work Future Work

_ _ Study how each storage component

Study how each storage component transforms locality characteristics transforms locality characteristics

_ _ Verify fidelity of various benchmarks

Verify fidelity of various benchmarks

_ _ Improve storage components and

Improve storage components and benchmarks benchmarks