This repository holds the simulation code and data analysis code for the Hyperbolic Caching Project. For more details, check out our paper.
To check out our Redis Prototype, see this repository.
I run the simulator in pypy and use NumPy for efficient memory and fast distribution sampling. Setting these up on Ubuntu LTS's can be a real joy.
These are roughly the commands I use to install these, setting up a virtualenv to run the simulator in.
$ sudo apt-get install -y pypy pypy-dev python-virtualenv
$ virtualenv --python=pypy ve
$ ve/bin/pip install python-dateutil cffi llist sortedcontainers psycopg2cffi
$ git clone https://bitbucket.org/pypy/numpy.git
$ cd numpy/
$ git fetch && git checkout pypy-2.6.0
$ ../ve/bin/pypy setup.py install
And then clone into this repository
$ git clone [email protected]:kantai/cost-cache.git
The simulator's code is primarily accessed through
simulation/main.py
The structure of this code is esoteric at best,
but no intentional obfuscation occurred.
To facilitate easy testing and changing of parameters, I eschewed the usage of a traditional CLI, opting instead to call functions directly as in:
$ python -c 'import main as m; m.hyper_v_hyper_class_mrc_hotclass()'
The data-files that ultimately became figures and results in the paper all have functions that will run their experiments directly.
If you'd like to 'peel back the covers' on these functions, what they
do is set up a Simulation
object which interacts with a workload
driver (usually referred to by d
in the code) and a caching
policy (useually referred to by p
in the code).
The function run_product
will perform some argument broadcasting to
construct a driver (passing it some parameters) and a policy (passing
it other parameters). The policies and drivers themselves are supplied
using names registered in main_globals.py
At various points in time, I thought it was a good idea to use the
simulator to drive experiments on a real cache with a real database
backend. Towards that end, I implemented subclasses of the Simulation
class (implemented in simulate.py
)-- RedisSim
in
instrument_redis.py
and BackedRedisSim
in the conspicuously named
expirement_db_redis.py
. When last I looked, these files worked, but
I didn't use them for any of the data collection in the paper, nor
have I used them particularly recently. However, I do use
setup_pgsql_sim
in main.py
to initialize the database that I used
in the Node simulations.
If you're interested in this, the function
run_redis_backed_spc_sim()
is going to be your friend.
Setting up node:
cd node-sim
curl -sL https://deb.nodesource.com/setup_4.x | sudo -E bash -
sudo apt-get install -y nodejs libpq-dev
npm i
I used IPTables because I hate running test environments on open ports on the SNS machines.
# IPTABLES configuration:
sudo iptables -I INPUT -p TCP --dport 3590 -j DROP
sudo iptables -I INPUT -p TCP --dport 3590 -s $REQUEST_MACHINE -j ACCEPT
sudo iptables -I INPUT -p TCP --dport 3590 -s localhost -j ACCEPT
You'll need to setup a postgres machine:
# IPTABLES configuration:
sudo iptables -I INPUT -p TCP --dport 5435 -j DROP
sudo iptables -I INPUT -p TCP --dport 5435 -s $NODE_MACHINE -j ACCEPT
sudo iptables -I INPUT -p TCP --dport 5435 -s localhost -j ACCEPT
sudo pg_createcluster -u $USER -p 5435 -d /disk/local/WHEREVER/ 9.3 $USER
pg_ctlcluster 9.3 $USER start
psql -p 5435 -h/tmp postgres -c "CREATE DATABASE $USER;"
echo "listen_addresses = '128.112.7.135'" >> /etc/postgresql/9.3/$USER/postgresql.conf
echo "host $USER $USER NODE_MACHINE_IP_ADDRESS/32 trust" >> /etc/postgresql/9.3/blanks/pg_hba.conf
pg_ctlcluster 9.3 $USER restart
And now this is the step where we fill the database using main.py
-- you'll need to have
installed NumPY in this virtualenv, per the instructions above.
cd cost-cache/simulation/
../../ve/bin/python -c "from main import setup_pgsql_sim; from main_globals import ZP1C_DRIVER; setup_pgsql_sim(ZP1C_DRIVER, 10**6)"
Now let's set up the request machine to run the websim -- it needs numpy and pycurl -- but can run on CPython, so just use Ubuntu's package manager:
sudo apt-get install pycurl numpy
And fire away from the request machine!
./run_multiple 0 zipfian_1_100k.csv 39000 3000
This script will attempt to spawn a node server, start redis, and then begin executing requests. The arguments specify the workload driver (numbered in websim.py), the output csv file, the number of requests per client, and then the number of simultaneous clients.
Getting the Django apps up and running is several degrees of tricky. In particular, filling the wikipedia database was unpleasant to say the least.
Let's start by setting up the app:
$ cd apps/django-wiki
$ tar -xzf django-wiki.tar.gz
$ virtualenv env && source env/bin/activate
$ cd wiki_project
$ pip install -r requirements.txt
$ pip install gunicorn
Before creating the app's tables and database, you should
edit wiki_project/settings.py
, then you can sync the db.
$ python manage.py syncdb
In simulation/workloads/wikipedia/
, there are a handful of scripts
used to load articles from a wikipedia XML dump.
The page dump we used was enwiki-20080103-pages-articles.xml.bz2
Download that dump and link it into the directory of extractor.py
which will use psycopg2
to fill your database. This script has a
bunch of hard-coded parameters for connecting to your database and
picking which table to use to fill the data -- you should change those
(hint, the tables get named when you run python manage.py syncdb
to
install the wiki project.) You should be editing the connection
parameters and the SQL insert string.
$ python extractor.py
Once you've done the above, getting the app running is pretty
easy. Just call gunicorn
with the app's WSGI module.
$ ../env/bin/gunicorn wiki_project.wsgi -b 0.0.0.0:8000 --workers 32
Follow the instructions in the app dir. (./django-middleware/apps/developer-ubuntu-com/README.md)
../env/bin/gunicorn developer_portal.wsgi -b 0.0.0.0:8213 --workers 24