precip 0.3

Pegasus Repeatable Experiments for the Cloud in Python - a simple module for setting up repeatable computer science experiments in the cloud

precip - Pegasus Repeatable Experiments for the Cloud in Python

Overview

Precip is a flexible exeperiment management API for running experiments
on clouds. Precip was developed for use on FutureGrid infrastructures
such as OpenStack, Eucalyptus (>=3.2), Nimbus, and at the same time
commercial clouds such as Amazon EC2. The API allows you to easily
provision resources, which you can then can run commands on and copy
files to/from subsets of instances identified by tags. The goal of the
API is to be flexible and simple to use in Python scripts to control
your experiments.

The API does not require any special images, which makes it easy to get
going. Any basic Linux image will work. More complex images can be used
if your experiment requires so, or you can use the experiment API to
run bootstrap scripts on the images to install/configure required
software.

A concept which simplfies interacting with the API is instance tagging.
When you start an instance, you can add arbitrary tags to it. The
instance also gets a set of default tags. API methods such as running a
remote command, or copying files, all use tags for specify which
instances you want to target.

Precip also handles ssh keys and security groups automatically. This is
done to make sure the experiment management is not interfering with
your existing cloud setup. The first time you use Precip, a directory
will be created called ~/.precip. Inside this directory, an unique
account id is stored for each machine along with a ssh keypair and this
keypair is used for accessing instances from multiple machines or
accounts at the same time. On clouds which supports it, the keypair is
automatically registered as 'precip_id', and a 'precip' security group
is created. If your experiement requires more ports to be open you can
use the cloud interface to add those ports to the precip security
group.

Precip is a fairly new API, and if you have questions or suggestions
for improvements, please contact [e-mail]

Installation

Prerequisites are the Paramiko and Boto Python modules. The Python
source package and RPMs are available at:
http://pegasus.isi.edu/static/precip/software/

API

provision(image_id, instance_type='m1.small', count=1, ebs_size=None,
tags=None, boot_timeout=600, boot_max_tries=3)
Provision a new instance. Note that this method starts the
provisioning cycle, but does not block for the instance to
finish booting. For blocking on instance creation/booting, see
wait()

Parameters:

+ image_id - the id of the image to instanciate
+ instance_type - the type of instance. This is infrastructure
specific, but usually follows the Amazon EC2 model with
m1.small, m1.large, and so on.
+ count - number of instances to create
+ ebs_size - the size in GB if you want to grow the root
filesystem EBS disk
+ tags - these are used to manipulate the instance later. Use
this to create logical groups of your instances.
+ boot_timeout - timeout in seconds for the instances to boot.
If the timeout is reached and the instance has been restarted
less than boot_max_tries, the instance will be restarted. If
the number of retries have been exhausted, an
ExperimentException is raised. The default is 600 seconds.
+ boot_max_tries - the number of times booting an instance is
allowed to be retried. The default value is 3.

wait(tags=[])
Barrier for all instances matching the tags argument. This
method will block until the instances have finish booting and
are accessible via their external hostnames.

Parameters:

+ tags - tags specifying the subset of instances to block on.
The default value is [] which means wait for all instances.

deprovision(tags)
Deprovisions (terminates) instances matching the tags argument

Parameters:

+ tags - tags specifying the subset of instances to deprovision.

list(tags)
Returns a list of details about the instances matching the tags.
The details include instance id, hostnames, and tags.

Parameters:

+ tags - tags specifying the subset of instances to give
information on. If you want details on all current instances,
use [].

Returns:

+ List of dictionaries, one for each instance.

get_public_hostnames(tags)
Provides a list of public hostnames for the instances matching
the tags. The public hostnames can be provided to other
instances in order to let the instances know about eachother.

Parameters:

+ tags - tags specifying the subset of instances.

Returns:

+ A list of public hostnames

get_private_hostnames(tags)
Provides a list of private hostnames for the instances matching
the tags. The private hostnames can be provided to other
instances in order to let the instances know about eachother.

Parameters:

+ tags - tags specifying the subset of instances.

Returns:

+ A list of private hostnames

get(tags, remote_path, local_path, user="root")
Transfers a file from a set of remote machines matching the
tags, and stores the file locally. If more than one instance
matches the tags, an instance id will be appended to the
local_path.

Parameters:

+ tags - these are used to manipulate the instance later. Use
this to create logical groups of your instances.
+ remote_path - the path of the file on the remote instance
+ local_path - the local path to tranfer to
+ user - remote user. If not specified, the default is 'root'

put(tags, local_path, remote_path, user="root")
Transfers a local file to a set of remote machines matching the
tags.

Parameters:

+ tags - these are used to manipulate the instance later. Use
this to create logical groups of your instances.
+ local_path - the local path to tranfer from
+ remote_path - the path on the remote instance to store the
file as
+ user - remote user. If not specified, the default is 'root'

run(tags, cmd, user="root", check_exit_code=True,
output_base_name=None)
Runs a command on the instances matches the tags. The commands
are run in series, on one instance after the other.

Parameters:

+ tags - these are used to manipulate the instance later. Use
this to create logical groups of your instances.
+ cmd - the command to run
+ user - remote user. If not specified, the default is 'root'.
If you need to run commands as another user, you will have to
make sure that user accepts the ssh key in ~/.precip/
+ check_exit_code - If set to True (default), commands returning
non-zero exit codes will result in a ExperimentException being
raised.
+ output_base_name - By default, the stdout and stderr of the
remote command is looked in the PRECIP stdout log. Giving a
base filename (.out and .err will be appended automatically)
will redirect the stdout and stderr to files instead.

Returns:

+ A list of lists, containing exit_code[], stdout[] and stderr[]
for the commands run

copy_and_run(tags, local_script, args=[], user="root",
check_exit_code=True)
Copies a script from the local machine to the remote instances
and executes the script. The script is run in series, on one
instance after the other.

Parameters:

+ tags - these are used to manipulate the instance later. Use
this to create logical groups of your instances.
+ local_script - the local script to run
+ args - arguments for the script
+ user - remote user. If not specified, the default is 'root'.
If you need to run commands as another user, you will have to
make sure that user accepts the ssh key in ~/.precip/
+ check_exit_code - If set to True (default), commands returning
non-zero exit codes will result in a ExperimentException being
raised.

Returns:

+ A list of lists, containing exit_code[], stdout[] and stderr[]
for the commands run

The basic methods above are standard across all the Cloud
infrastructures. What is different is the constructors as each
infrastructure handles initialization a little bit different. For
example, to create a new OpenStack using the EC2_* environment provided
automatically by FutureGrid:

exp = OpenStackExperiment(
os.environ['EC2_URL'],
os.environ['EC2_ACCESS_KEY'],
os.environ['EC2_SECRET_KEY'])


i For Amazon EC2, you have to specify region, endpoint, and
access/secret keys. Note that it is not required to use environment
variables for your credentials, but seperating the crenditals from the
code prevents the credentials from being check in to source control
systems.

exp = EC2Experiment(
"us-west-2c",
"ec2.us-west-2.amazonaws.com",
os.environ['AMAZON_EC2_ACCESS_KEY'],
os.environ['AMAZON_EC2_SECRET_KEY'])


Examples

Hello World


#!/usr/bin/python

import os
import time
from pprint import pprint

from precip import *

exp = None

# Use try/except liberally in your experiments - the api is set up to
# raise ExperimentException on most errors
try:

# Create a new OpenStack based experiment. In this case we pick
# up endpoints and access/secret cloud keys from the environment
# as exposing those is the common setup on FutureGrid
exp = OpenStackExperiment(
os.environ['EC2_URL'],
os.environ['EC2_ACCESS_KEY'],
os.environ['EC2_SECRET_KEY'])

# Provision an instance based on the ami-0000004c. Note that tags are
# used throughout the api to identify and manipulate instances. You
# can give an instance an arbitrary number of tags.
exp.provision("ami-0000004c", tags=["test1"], count=1)

# Wait for all instances to boot and become accessible. The provision
# method only starts the provisioning, and can be used to start a large
# number of instances at the same time. The wait method provides a
# barrier to when it is safe to start the actual experiment.
exp.wait()

# Print out the details of the instance. The details include instance id,
# private and public hostnames, and tags both defined by you and some
# added by the api
pprint(exp.list())

# Run a command on the instances having the "test1" tag. In this case we
# only have one instance, but if you had multiple instances with that
# tag, the command would run on each one.
exp.run(["test1"], "echo 'Hello world from a experiment instance'")

except ExperimentException as e:
# This is the default exception for most errors in the api
print "ERROR: %s" % e

finally:
# Be sure to always deprovision the instances we have started. Putting
# the deprovision call under finally: make the deprovisioning happening
# even in the case of failure.
if exp is not None:
exp.deprovision()


Resources from mulitple infrastructures


#!/usr/bin/python

import os
import time

from precip import *

ec2 = None
fg = None

try:

# This example show how to run an experiment between Amazon EC2
# and an OpenStack resource on FutureGrid. The setup is pretty
# similar to the HelloWorld example, except that we now have to
# experiment to handle. The first step is to get the experiments
# initialized. Note that it is not required to use environment
# variables for your credentials, but seperating the crenditals
# from the code prevents the credentials from being check in to
# source control systems.

ec2 = EC2Experiment(
"us-west-2c",
"ec2.us-west-2.amazonaws.com",
os.environ['AMAZON_EC2_ACCESS_KEY'],
os.environ['AMAZON_EC2_SECRET_KEY'])

fg = OpenStackExperiment(
os.environ['EC2_URL'],
os.environ['EC2_ACCESS_KEY'],
os.environ['EC2_SECRET_KEY'])

# Next we provision two instances, one on Amazon EC2 and one of
# FutureGrid
ec2.provision("ami-8a1e92ba", tags=["id=ec2_1"])
fg.provision("ami-0000004c", tags=["id=fg_1"])

# Wait for all instances to boot and become accessible. The provision
# method only starts the provisioning, and can be used to start a large
# number of instances at the same time. The wait method provides a
# barrier to when it is safe to start the actual experiment.
ec2.wait([])
fg.wait([])

# Run commands on the remote instances
ec2.run([], "echo 'Hello world Amazon EC2'")
fg.run([], "echo 'Hello world FutureGrid OpenStack'")

except ExperimentException as e:
# This is the default exception for most errors in the api
print "ERROR: %s" % e
raise e
finally:
# Be sure to always deprovision the instances we have started. Putting
# the deprovision call under finally: make the deprovisioning happening
# even in the case of failure.
if ec2 is not None:
ec2.deprovision([])
if fg is not None:
fg.deprovision([])


Setting up a Condor pool and run a Pegasus workflow

This is a more complex example in which a small Condor pool is set up
and then a Pegasus workflow is run and benchmarked. The Precip script
is similar to what we have seen before, but it has two groups of
instances, one master acting as the Condor central manager, and a set
of Condor worker nodes.

#!/usr/bin/python

import os
import time

from precip import *

try:

# This experiment is targeted to run on OpenStack
exp = OpenStackExperiment(
os.environ['OPENSTACK_URL'],
os.environ['OPENSTACK_ACCESS_KEY'],
os.environ['OPENSTACK_SECRET_KEY'])

# We need a master Condor node and a set of workers
exp.provision("ami-0000004c", tags=["master"],
instance_type="m1.large")
exp.provision("ami-0000004c", tags=["compute"],
instance_type="m1.large", count=2)
exp.wait()

# The workers need to know what the private hostname of the master is
master_priv_addr = exp.get_private_hostnames(["master"])[0]

# Bootstrap the instances. This includes installing Condor and Pegasus,
# downloading and settup the workflow.
exp.copy_and_run(["master"], "./bootstrap.sh")
exp.copy_and_run(["compute"], "./bootstrap.sh", args=[master_priv_addr])

# Give the workers some time to register with the Condor central
# manager
time.sleep(60)

# Make sure Condor came up correctly
exp.run(["master"], "condor_status")

# Run the workflow
exp.run(["master"], "cd ~/montage && ./run-montage", user="wf")

# At this point, in a real experiment, you could for example provision
# more resources and run the workflow again, or run the workflow with
# different parameters/settings.

except ExperimentException as e:
print "ERROR: %s" % e
finally:
# always want to clean up all the instances we have started
exp.deprovision([])


We also need a bootstrap.sh which sets up the instances:

#!/bin/bash

# This script bootstraps a basic RHEL6 instance to be have working
# Condor and Pegasus installs. The script takes one optional
# argument which is the address of the master instance (central
# manager in Condor terminology). If the argument is not given,
# the script sets up the instance to be the master.

MASTER_ADDR=$1

# for images with condor already installed, stop condor
/etc/init.d/condor stop >/dev/null 2>&1 || /bin/true

# correct clock is important for most projects
yum -q -y install ntpdate
/etc/init.d/ntpdate start

# Add the EPEL repository
wget -nv http://mirror.utexas.edu/epel/6/x86_64/epel-release-6-7.noarch.rpm
rpm -Uh epel-release-*.noarch.rpm

# Add the Condor repository
cat >/etc/yum.repos.d/condor.repo <<EOF
[condor-stable]
name=Condor Stable RPM Repository for Redhat Enterprise Linux 6
baseurl=http://www.cs.wisc.edu/condor/yum/stable/rhel6
enabled=1
gpgcheck=0
EOF

# Add the Pegasus repository
cat >/etc/yum.repos.d/pegasus.repo <<EOF
[Pegasus]
name=Pegasus
baseurl=http://download.pegasus.isi.edu/wms/download/rhel/6/x86_64
gpgcheck=0
enabled=1
priority=50
EOF

# Install required software
yum -q -y clean all
yum -q -y install gcc gcc-g++ gcc-gfortran make gawk bison diffutils \
java-1.7.0-openjdk.x86_64 \
java-1.7.0-openjdk-devel.x86_64 \
ganglia-gmond condor pegasus

# Clear the Condor local config file - we use config.d instead
cat /dev/null >/etc/condor/condor_config.local

# Common Condor config between master and workers
cat >/etc/condor/config.d/50-main.conf <<EOF

CONDOR_HOST = $MASTER_ADDR

FILESYSTEM_DOMAIN = \$(FULL_HOSTNAME)
TRUST_UID_DOMAIN = True

DAEMON_LIST = MASTER, STARTD

# security
ALLOW_WRITE = 10.*
ALLOW_READ = \$(ALLOW_WRITE)

# default policy
START = True
SUSPEND = False
CONTINUE = True
PREEMPT = False
KILL = False

EOF

# Master gets extra packages/configs
if [ "x$MASTER_ADDR" = "x" ]; then
yum -q -y install ganglia-gmetad ganglia-web

cat >/etc/condor/config.d/90-master.conf <<EOF
CONDOR_HOST = \$(FULL_HOSTNAME)
DAEMON_LIST = MASTER, COLLECTOR, NEGOTIATOR, SCHEDD
EOF
fi

# Restarting daemons
/etc/init.d/condor start

# User to run the workflows as, and allow the experiment management
# ssh key to authenticate
adduser wf
mkdir -p ~wf/.ssh
cp ~/.ssh/authorized_keys ~wf/.ssh/
chown -R wf: ~wf/.ssh

# Master is the submit host, so deploy our workflow on it
if [ "x$MASTER_ADDR" = "x" ]; then
# install the workflow tarball and wait script
cd ~wf
wget -q http://pegasus.isi.edu/static/precip/wf-experiment/montage.tar.gz
tar xzf montage.tar.gz
chown -R wf: montage*
fi