Polca

Use LearnLib to learn automata models for cache replacement policies.

Support for several simulated replacement policies.

Binding with real hardware relies on CacheQuery. (add link)

Install

mvn install

We use LearnLib 0.14.0: https://learnlib.github.io/learnlib/maven-site/0.14.0/apidocs/

Run

Help menu:

 usage: Polca
 -b,--binary <arg>        path to proxy for 'hw' policy
 -d,--depth <arg>         max_depth for membership queries (default: 1)
 -h,--help                show this help message
 -hit_ratio <arg>         ratio of hits to consider a HIT (default: 0.8)
 -l,--learner <arg>       learning algorithm lstar|kv|mp|rs|dhc|dt|ttt
                          (default: 'kv')
 -m,--max_size <arg>      maximum number of states of SUL
 -miss_ratio <arg>        ratio of misses to consider a MISS (default:
                          0.2)
 -no_cache                don't use cache for membership queries
 -o,--output <arg>        write learnt .dot model into output file
 -p,--policy <arg>        simulator cache policy:
                          fifo|lru|plru|lip|plip|mru|srriphp|srripfp|s
                          kyl2|skyl3|hw (default: 'fifo')
 -prefix <arg>            prefix before every query, used to fill cache
                          (default: '@')
 -r,--repetitions <arg>   number of measurements by cachequery (default:
                          100)
 -r_bound <arg>           bound on queries for equivalence, set to 0 for
                          unbounded (default: 1000)
 -r_len <arg>             expected length of random word (r_min + r_len)
                          (default: 30)
 -r_min <arg>             minimal length of random word (default: 10)
 -r_rand <arg>            TODO: select custom random generator
 -random                  use random wp-method as equivalence query
 -s,--silent              remove stdout info
 -temp                    write partial model into '.model.tmp' file
 -verbose                 output verbose information
 -votes <arg>             number of votes for deciding result (default: 1)
 -w,--ways <arg>          cache associativity (default: 4)

Example for learning LRU assoc=4 from simulator:

./polca.sh -w 4 -p lru -verbose

Example for learning L1 in Haswell machine:

./polca.sh -w 8 -p hw -b \\\"ssh -t pepe@haswell ~/cachequery/cachequery.py -c ~/cachequery/cachequery.ini -i -l l1\\\" -prefix \\\"@ @\\\" -verbose

Docker

To build the project and all dependencies in a docker container run:

docker build -t polca .

Make tmp/ directory writeable (i.e. chmod o+w tmp/) to others, and run the following to get a command line into the container:

sudo docker run -ti -v $(pwd)/tmp:/home/user/polca/tmp polca

The tmp/ directory is mounted inside the container in order to share output files with the host.

In order to verify that everything works as expected run:

user@xxx:~/polca$ ./polca.sh --help

Test 1 - Learn PLRU assoc=4 from simulator

$ ./polca.sh -w 4 -p plru -o tmp/plru.dot

In order to visualize the file from the host machine, run:

$ dot -Tpng tmp/plru.dot | feh -

Or, if dot or feh are not installed in the host system, convert the dot file into a PNG image inside the container:

dot -Tpng tmp/plru.dot -o tmp/plru.png

And open it with any tool of your choice.

Feel free to modify the associativty or the policy under learning. All the learned policies from a simulator are available at models/simul/.

Test 2 - Synthesize exaplanation for LRU

First of all we need to convert the automata model into a Sketch file, for this we run:

$ bash scripts/dot_to_constraints.sh models/simul/lru_4.dot 4 | tee tmp/lru4.sk

Compare (e.g. via diff) the resulting file with ~/polca/scripts/sketch/lru4.sk, and run:

$ cd scripts/sketch/
$ sketch lru4.sk | tee ../../tmp/lru.out

The resulting file is an explanation for the LRU policy.

Directory scripts/sketch/ contains all the Sketch files with the corresponding automata constraints; scripts/sketch/output/ contains all the explanations/solutions (as *.out files) and a manually cleaned version of them (as *.clean files).

Hardware Examples

CPU	Level	Ways	States	Description	Link to model	Prefix	HW mem queries	HW eq queries
Haswell i7-4790	L1	8	128	Tree-PLRU	dot	@ @	3584	36158
Haswell i7-4790	L2	8	128	Tree-PLRU	dot	@	3584	36158
Skylake i5-6500	L1	8	128	Tree-PLRU	dot	@	3584	36158
Skylake i5-6500	L2	4	160	QLRU variant	dot	d c b a @	3525	42225
Skylake i5-6500	L3	4 (with CAT)	175	QLRU variant	dot	@	3804	45117
KabyLake R i7-8550U	L1	8	128	Tree-PLRU	dot	@	3584	36158
KabyLake R i7-8550U	L2	4	160	QLRU variant	dot	d c b a @	3525	42225
KabyLake R i7-8550U	L3	4 (with CAT)	175	QLRU variant	dot	@	3804	45117

We include the learnt models (.dot files) under the models/ directory.

Prefixes are used to fill the cache with initial content and put it to the same control state.

Simulator Examples

We also include learnt models for all the policies supported by simulator (see models/simul/), with associativity 4:

Policy	States (assoc=n)	States (assoc=4)	Link to model	Learning time
FIFO	n	4	dot	0.081s
LRU	n!	24	dot	0.307s
PLRU	2^(n-1)	8	dot	0.155s
LIP	n!	24	dot	0.275s
MRU	2^n - 2	14	dot	0.179s
SRRIPFP-4	4^n	256	dot	7.618s
SRRIPHP-4	?	178	dot	2.097s

Issues

• No support for TTY: sudo's prompt doesn't work trough Polca, need to use root or NOPASSWD in /etc/sudoers.
• ...