def create_dict(self, name, keys):
worker_key_pairs = sorted([(hash(key) % self.size, key) for key in keys])
for machine, local_group in itertools.groupby(sorted(worker_key_pairs), lambda x: x[0]):
local_dict = {}
for machine, key in local_group:
local_dict[key] = {}
mr_lib.write_pickle(local_dict, name)
self.send(name, machine)
def map(self, filename, input_dict, mapper_params):
# distribute the file defining the mapper functions through the cluster,
# as well as the corresponding parameters.
self.distribute_public(filename)
mr_lib.write_pickle(mapper_params, "mapper_params.mr")
self.distribute_public("mapper_params.mr")
# Launch the map phase, and wait until it's done.
self.exec_public("python map.py " + filename + " " + input_dict)
self.wait_until_task_done("map_done")
def mr(self, filename, mapper_params, reducer_params, input_dict, output_dict, output_field):
# distribute the file defining the mapper and reducer functions through the cluster,
# as well as the corresponding parameters
self.distribute_public(filename)
mr_lib.write_pickle([mapper_params, reducer_params], "params.mr")
self.distribute_public("params.mr")
# Launch the map phase, with combining, and wait until it's done
self.exec_public("python map_combine.py " + filename + " " + input_dict)
self.wait_until_task_done("map_combine_done")
# Launch the reduce phase, and wait until it's done
self.exec_public("python reduce.py " + filename + " " + output_dict + " " + output_field)
self.wait_until_task_done("reduce_done")
def __init__(self, n):
# set the size attribute
self.size = n
# set up an EC2 connection, and grab an Ubuntu 8.04 image (http://alestic.com)
connection = boto.connect_ec2()
image = connection.get_image("ami-1c5db975")
# create a keypair to use with the image, save to disk, and set permissions
# so ssh will be happy
self.keypair = connection.create_key_pair("mr_keypair")
mr_lib.write_file(self.keypair.material, mr_lib.mr_keypair_filename())
os.system("chmod 600 " + mr_lib.mr_keypair_filename())
# tell EC2 to start the instances running, set the self.workers attribute to the
# corresponding reservation, and wait for all the workers to start running
self.workers = image.run(n, n, "mr_keypair")
for instance in self.workers.instances:
instance.update()
while instance.state != u"running":
instance.update()
time.sleep(5)
# Delay before we start distributing files, so all instances are running properly.
time.sleep(10)
# distribute a list of all the private ip addresses
private_ip_list = [instance.private_dns_name for instances in self.workers.instances]
mr_lib.write_pickle(private_ip_list, "cluster_description.mr")
self.distribute_public("cluster_description.mr")
for j in xrange(n):
mr_lib.write_pickle([j, self.workers.instances[j].private_dns_name], "my_details.mr")
self.send("my_details.mr", j)
# distribute the files necessary to run map and mapreduce jobs
self.distribute_public("map.py")
self.distribute_public("map_combine.py")
self.distribute_public("reduce.py")
self.distribute_public("mr_lib.py")
# Distribute the ssh keypairs and config file
for instance in self.workers.instances:
mr_lib.scp(mr_lib.mr_keypair_filename(), "root@" + instance.public_dns_name + ":.ssh/id_rsa-mr_keypair")
mr_lib.ssh("root@" + instance.public_dns_name, "chmod 600 /root/.ssh/id_rsa-mr_keypair")
mr_lib.scp(os.environ.get("HOME") + "/.ssh/config", "root@" + instance.public_dns_name + ":.ssh/config")
# map.py
#
# Part of the mr.py library. Run on worker machines by mr.py, when a map
# job only is called.
import itertools, mr_lib, sys
mr_lib.set_flag("map_done",False) # map phase on this worker not yet done
filename,input_dict = sys.argv[1:]
exec("from "+filename+" import mapper") # import map job
i = mr_lib.read_pickle(input_dict) # Read the input dictionary
mapper_params = mr_lib.read_pickle("mapper_params.mr") # Get the parameters for the mapper
for k in i.keys(): i[k] = mapper(k,i[k],mapper_params) # Run the mapper
mr_lib.write_pickle(i,input_dict) # Write the input dictionary back out
mr_lib.set_flag("map_done",True) # map phase is done
filename,output_dict,output_field = sys.argv[1:]
# Get the parameters for the MapReduce job
mapper_params,reducer_params = mr_lib.read_pickle("params.mr")
# import the reducer
module = filename[:-3]
exec("from "+module+" import reducer")
# read in all the intermediate data
intermediate = []
for machine in xrange(len(ip)):
name = "inter.dict."+str(machine)+"."+str(my_number)
if os.path.exists(name): intermediate.extend(mr_lib.read_pickle(name))
groups = {}
for key, group in itertools.groupby(sorted(intermediate),
lambda x: x[0]):
groups[key] = list([y for x, y in group])
result = [reducer(inter_key,groups[inter_key],reducer_params) for inter_key in groups]
# load the existing output dictionary, and modify it to include the results
# obtained in the reduce step, saving the resulting file.
o = mr_lib.read_pickle(output_dict)
for key, value in result: o[key][output_field] = value
mr_lib.write_pickle(o,output_dict)
# Sets a flag, visible to the client, saying that the reduce phase on this worker
# is done.
mr_lib.set_flag("reduce_done",True)
