jcipar / stack-distance

Utility to simulate cache behavior with Mattson's Stack Algorithm.

stack-distance

Utility to simulate cache behavior with Mattson's Stack Algorithm. This is something I've had to rewrite a number of times for various projects. I finally decided to write it and put it somewhere where I, and others, could reuse it.

Mattson's Stack Algorithm is used to estimate the cache miss rate for a given workload, with any sized cache. Given a trace of memory/disk/object accesses, it can be used to find a cache size that would be appropriate for that workload.

It uses the assumption that most real-world caches behave similarly- enough to an LRU that the LRU miss rate is "good enough". The algorithm simulates an LRU cache. For each memory access, it gives the LRU position of that access. By giving it a stream of object IDs, it will print out a stream of LRU positions. A cumulative histogram of these LRU positions is a "miss ratio curve": a plot showing, for each cache size, what the cache hit rate would have been for that cache size.

Using the utility

The stack-distance program takes object ids as arbitrary strings on standard input. For each line of input, it prints an integer LRU position. Here's an example. Object IDs are letters, LRU positions are numbers, and comments begin with #.

a
9223372036854775807  # INT64_MAX is used for new objects.
b
9223372036854775807
c
9223372036854775807
a
2  # Two unique references since the last reference to a.
a
0  # A was the last reference, so it now has a distance of 0.
b
2  # Even though a has appeared twice, there are only 2 *unique* references
   # since the last time we saw b.
a
1
c
2

The output of this utility can be passed to sort -n | uniq -c to generate a histogram.

Accounting for Object size

The --weight parameter can be used to simulate objects of different sizes. When this parameter is used, every input is expected to be a string identifier followed by an integer size. The outputs are weighted by size.

a 10  # Add an object a, size 10
9223372036854775807  # INT64_MAX is used for new objects.
b 30  # Add an object b, size 30
9223372036854775807
c 15  # Add an object c, size 15
9223372036854775807
a 10  # SIZE IS REQUIRED ON EVERY REFERENCE
45    # Total weight of 45 since the last access to a 
a 10  
0     # A was the last reference, so it now has a distance of 0.
b 30  
25  # Even though a has appeared twice, there are only 2 *unique* references
    # since the last time we saw b.
a
30
c
40

Computing miss-ratio-curve for a large social network graph

gunzip --stdout twitter_rv.tar.gz  | 
	pv -s 26172280241 -p -t -e -r -a | 
	stack-distance | 
	sort -n | 
	uniq -c > twitter-mrc.txt