ceph/qa/tasks/radosbenchsweep.py

223 lines
7.5 KiB
Python
Raw Normal View History

"""
Rados benchmarking sweep
"""
import contextlib
import logging
import re
from io import BytesIO
from itertools import product
from teuthology.orchestra import run
from teuthology import misc as teuthology
log = logging.getLogger(__name__)
@contextlib.contextmanager
def task(ctx, config):
"""
Execute a radosbench parameter sweep
Puts radosbench in a loop, taking values from the given config at each
iteration. If given, the min and max values below create a range, e.g.
min_replicas=1 and max_replicas=3 implies executing with 1-3 replicas.
Parameters:
clients: [client list]
time: seconds to run (default=120)
sizes: [list of object sizes] (default=[4M])
mode: <write|read|seq> (default=write)
repetitions: execute the same configuration multiple times (default=1)
min_num_replicas: minimum number of replicas to use (default = 3)
max_num_replicas: maximum number of replicas to use (default = 3)
min_num_osds: the minimum number of OSDs in a pool (default=all)
max_num_osds: the maximum number of OSDs in a pool (default=all)
file: name of CSV-formatted output file (default='radosbench.csv')
columns: columns to include (default=all)
- rep: execution number (takes values from 'repetitions')
- num_osd: number of osds for pool
- num_replica: number of replicas
- avg_throughput: throughput
- avg_latency: latency
- stdev_throughput:
- stdev_latency:
Example:
- radsobenchsweep:
columns: [rep, num_osd, num_replica, avg_throughput, stdev_throughput]
"""
log.info('Beginning radosbenchsweep...')
assert isinstance(config, dict), 'expecting dictionary for configuration'
# get and validate config values
# {
# only one client supported for now
if len(config.get('clients', [])) != 1:
raise Exception("Only one client can be specified")
# only write mode
if config.get('mode', 'write') != 'write':
raise Exception("Only 'write' mode supported for now.")
# OSDs
total_osds_in_cluster = teuthology.num_instances_of_type(ctx.cluster, 'osd')
min_num_osds = config.get('min_num_osds', total_osds_in_cluster)
max_num_osds = config.get('max_num_osds', total_osds_in_cluster)
if max_num_osds > total_osds_in_cluster:
raise Exception('max_num_osds cannot be greater than total in cluster')
if min_num_osds < 1:
raise Exception('min_num_osds cannot be less than 1')
if min_num_osds > max_num_osds:
raise Exception('min_num_osds cannot be greater than max_num_osd')
osds = range(0, (total_osds_in_cluster + 1))
# replicas
min_num_replicas = config.get('min_num_replicas', 3)
max_num_replicas = config.get('max_num_replicas', 3)
if min_num_replicas < 1:
raise Exception('min_num_replicas cannot be less than 1')
if min_num_replicas > max_num_replicas:
raise Exception('min_num_replicas cannot be greater than max_replicas')
if max_num_replicas > max_num_osds:
raise Exception('max_num_replicas cannot be greater than max_num_osds')
replicas = range(min_num_replicas, (max_num_replicas + 1))
# object size
sizes = config.get('size', [4 << 20])
# repetitions
reps = range(config.get('repetitions', 1))
# file
fname = config.get('file', 'radosbench.csv')
f = open('{}/{}'.format(ctx.archive, fname), 'w')
f.write(get_csv_header(config) + '\n')
# }
# set default pools size=1 to avoid 'unhealthy' issues
ctx.manager.set_pool_property('data', 'size', 1)
ctx.manager.set_pool_property('metadata', 'size', 1)
ctx.manager.set_pool_property('rbd', 'size', 1)
current_osds_out = 0
# sweep through all parameters
for osds_out, size, replica, rep in product(osds, sizes, replicas, reps):
osds_in = total_osds_in_cluster - osds_out
if osds_in == 0:
# we're done
break
if current_osds_out != osds_out:
# take an osd out
ctx.manager.raw_cluster_cmd(
'osd', 'reweight', str(osds_out-1), '0.0')
wait_until_healthy(ctx, config)
current_osds_out = osds_out
if osds_in not in range(min_num_osds, (max_num_osds + 1)):
# no need to execute with a number of osds that wasn't requested
continue
if osds_in < replica:
# cannot execute with more replicas than available osds
continue
run_radosbench(ctx, config, f, osds_in, size, replica, rep)
f.close()
yield
def get_csv_header(conf):
all_columns = [
'rep', 'num_osd', 'num_replica', 'avg_throughput',
'avg_latency', 'stdev_throughput', 'stdev_latency'
]
given_columns = conf.get('columns', None)
if given_columns and len(given_columns) != 0:
for column in given_columns:
if column not in all_columns:
raise Exception('Unknown column ' + column)
return ','.join(conf['columns'])
else:
conf['columns'] = all_columns
return ','.join(all_columns)
def run_radosbench(ctx, config, f, num_osds, size, replica, rep):
pool = ctx.manager.create_pool_with_unique_name()
ctx.manager.set_pool_property(pool, 'size', replica)
wait_until_healthy(ctx, config)
log.info('Executing with parameters: ')
log.info(' num_osd =' + str(num_osds))
log.info(' size =' + str(size))
log.info(' num_replicas =' + str(replica))
log.info(' repetition =' + str(rep))
for role in config.get('clients', ['client.0']):
assert isinstance(role, str)
PREFIX = 'client.'
assert role.startswith(PREFIX)
id_ = role[len(PREFIX):]
(remote,) = ctx.cluster.only(role).remotes.keys()
proc = remote.run(
args=[
'adjust-ulimits',
'ceph-coverage',
'{}/archive/coverage'.format(teuthology.get_testdir(ctx)),
'rados',
'--no-log-to-stderr',
'--name', role,
'-b', str(size),
'-p', pool,
'bench', str(config.get('time', 120)), 'write',
],
logger=log.getChild('radosbench.{id}'.format(id=id_)),
stdin=run.PIPE,
stdout=BytesIO(),
wait=False
)
# parse output to get summary and format it as CSV
proc.wait()
out = proc.stdout.getvalue()
all_values = {
'stdev_throughput': re.sub(r'Stddev Bandwidth: ', '', re.search(
r'Stddev Bandwidth:.*', out).group(0)),
'stdev_latency': re.sub(r'Stddev Latency: ', '', re.search(
r'Stddev Latency:.*', out).group(0)),
'avg_throughput': re.sub(r'Bandwidth \(MB/sec\): ', '', re.search(
r'Bandwidth \(MB/sec\):.*', out).group(0)),
'avg_latency': re.sub(r'Average Latency: ', '', re.search(
r'Average Latency:.*', out).group(0)),
'rep': str(rep),
'num_osd': str(num_osds),
'num_replica': str(replica)
}
values_to_write = []
for column in config['columns']:
values_to_write.extend([all_values[column]])
f.write(','.join(values_to_write) + '\n')
ctx.manager.remove_pool(pool)
def wait_until_healthy(ctx, config):
first_mon = teuthology.get_first_mon(ctx, config)
(mon_remote,) = ctx.cluster.only(first_mon).remotes.keys()
teuthology.wait_until_healthy(ctx, mon_remote)