def onlinecpu():
import numa
o_cpus = []
for node in range(0, numa.get_max_node() + 1):
for cpu in sorted(numa.node_to_cpus(node)):
o_cpus.append(cpu)
return o_cpus
def check_numa():
try:
import numa
except:
return
if not numa.available():
return
if numa.get_max_node() > 0 and len(numa.get_run_on_node_mask()) > 1:
print("Warning: NUMA settings may be suboptimal!", file=sys.stderr)
def setNuma():
"""Without subapAllocation... whole rmx on each numa node"""
d = darc.Control()
nthr = d.Get("ncamThreads").sum()
nnodes = numa.get_max_node() + 1
#specify which numa nodes the threads are closest to:
threadToNuma = numpy.zeros(nthr, numpy.int32) #all node 0 for now...!
rmx = d.Get("gainReconmxT")
d.Set("threadToNuma", threadToNuma)
for i in range(nnodes):
d.SetNuma("gainReconmxT%d" % i, rmx, i)
def load_numa():
""" Load information about core numbers and numa patterns """
if not numa.available():
raise Exception('Numa detection not available')
max_node = numa.get_max_node()
nodes = {}
for i in range(max_node + 1):
nodes[i] = list(numa.node_to_cpus(i))
return nodes
def coremap():
try:
import numa
except ImportError:
print('This script requires the libnuma python bindings')
raise RuntimeError("Numa not available")
if not numa.available():
raise RuntimeError("Numa not available")
node_to_core = {
int(i): deque([int(k) for k in numa.node_to_cpus(i)])
for i in range(numa.get_max_node() + 1)
}
total_core = max(itertools.chain(*node_to_core.values())) + 1
return node_to_core, total_core
def main():
parser = argparse.ArgumentParser()
parser.add_argument("-membw", default="./membw/membw")
parser.add_argument("-bandwidth", type=int, default=9999)
parser.add_argument("-operation", default="nt-write")
args = parser.parse_args()
max_nid = numa.get_max_node()
if max_nid != 1:
print("This tool requires two sockets at least")
sys.exit()
cpus = numa.node_to_cpus(max_nid)
for cpuid in cpus:
cmd = "%s -c %d -b %d --%s &" % (args.membw, cpuid, args.bandwidth, args.operation)
print(cmd)
os.system(cmd)
def _configure_numa(self):
self._numa_available = \
numa.available() and which('numactl') is not None
if not self._numa_available:
return
num_numa_nodes = numa.get_max_node() + 1
self._numa_cpu_map = {}
num_gpus = len(self._gpu_ids)
# Calculate how many CPUs to allocate for each GPU. Ensure this number
# is a power of 2.
num_cpus = 0
for i in range(num_numa_nodes):
num_cpus += len(numa.node_to_cpus(i))
num_cpus_per_gpu = min(MAX_CPUS_PER_GPU, max(num_cpus // num_gpus, 1))
num_cpus_per_gpu = pow(2, round(math.log(num_cpus_per_gpu, 2)))
# Find blocks of contiguous CPUs.
contiguous_blocks = []
for i in range(num_numa_nodes):
cpus = sorted(numa.node_to_cpus(i))
contiguous_block = [cpus[0]]
for j in range(1, len(cpus)):
if (cpus[j] - cpus[j - 1] == 1
and len(contiguous_block) < num_cpus_per_gpu):
contiguous_block.append(cpus[j])
else:
contiguous_blocks.append(
(contiguous_block, len(contiguous_block)))
contiguous_block = [cpus[j]]
if len(contiguous_block) > 0:
contiguous_blocks.append(
(contiguous_block, len(contiguous_block)))
contiguous_blocks.sort(key=lambda x: x[-1], reverse=True)
# Assign CPUs to GPUs.
block_idx = 0
for i in range(num_gpus):
self._numa_cpu_map[i] = []
while len(self._numa_cpu_map[i]) < num_cpus_per_gpu:
self._numa_cpu_map[i] += contiguous_blocks[block_idx][0]
block_idx = (block_idx + 1) % len(contiguous_blocks)
self._logger.info('GPU {gpu} assigned CPUs {cpus}'.format(
gpu=i, cpus=str(self._numa_cpu_map[i])))
def _lazy_cpu_and_mem_set_init(self):
# Implicitly assume lock is already held
if len(self._numa_nodes) != 0:
# Init already happened
return
if (self._available_cpu_ids is None or self._cpus_per_job is None
or self._use_memset_of_nearest_node is None):
raise Exception('Cannot do init. One or more params were None')
import numa
if not numa.available():
raise Exception('NUMA not available')
numa_nodes = list(range(0, numa.get_max_node() + 1))
cpu_count = 0
for numa_node in numa_nodes:
cpus = numa.node_to_cpus(numa_node)
for cpu_id in cpus:
if cpu_id in self._available_cpu_ids:
try:
self._numa_nodes[numa_node].add(cpu_id)
except KeyError:
self._numa_nodes[numa_node] = set()
self._numa_nodes[numa_node].add(cpu_id)
try:
self._numa_node_pool[numa_node].add(cpu_id)
except KeyError:
self._numa_node_pool[numa_node] = set()
self._numa_node_pool[numa_node].add(cpu_id)
_logger.info(
'Putting CPU {} in NUMA node {} in resource pool'.
format(cpu_id, numa_node))
cpu_count += 1
else:
_logger.info(
'CPU {} in NUMA node {} is NOT IN resource pool'.
format(cpu_id, numa_node))
if cpu_count == 0:
raise Exception('Found no available CPUs')
if cpu_count != len(self._available_cpu_ids):
raise Exception(
'Mismatch between provided available CPU ids and what was found on system'
)
assert len(self._numa_node_pool) == len(self._numa_nodes)
def setNuma2():
"""With subapAllocation: partial rmx on each numa node"""
d = darc.Control()
nthr = d.Get("ncamThreads").sum()
nnodes = numa.get_max_node() + 1
#specify which numa nodes the threads are closest to:
threadToNuma = numpy.zeros(nthr, numpy.int32) #all node 0 for now...!
sf = d.Get("subapFlag")
nsub = sf.size
sa = numpy.zeros(nsub, numpy.int32)
#divide all threads equally...
sa[:] = nthr - 1
start = 0
for i in range(nthr):
end = start + nsub / nthr
sa[start:end] = i
start = end
print numpy.reshape(sa, (7, 7))
print numpy.reshape(sf, (7, 7))
d.Set("threadToNuma", threadToNuma, swap=1)
rmx = d.Get("gainReconmxT") #nslope,nact
thrSubapCnt = numpy.zeros(nthr, numpy.int32)
rmxPart = []
for i in range(nthr):
rmxPart.append([])
indx = 0
for i in range(nsub):
if sf[i]:
thrSubapCnt[sa[i]] += 1
rmxPart[sa[i]].append(rmx[indx])
rmxPart[sa[i]].append(rmx[indx + 1])
indx += 2
for i in range(nthr):
r = numpy.array(rmxPart[i])
print "Thread %d rmx shape %s, dtype %s" % (i, str(r.shape), r.dtype)
d.SetNuma("gainReconmxT%d" % i, r, int(threadToNuma[i]), swap=1)
d.Set("subapAllocation", sa, swap=1)
def __init__(self):
gr.top_block.__init__(self, "ATA New SNAP X-Engine")
##################################################
# Variables
##################################################
self.starting_channel = starting_channel = clparam_starting_channel
self.num_channels = num_channels = clparam_num_channels
self.output_file = output_file = clparam_output_directory + '/casa_2021_jan_04_sync_v3_xeng'
self.ending_channel = ending_channel = starting_channel + num_channels - 1
##################################################
# Blocks
##################################################
self.clenabled_clXEngine_0 = clenabled.clXEngine(
1, 2, 0, 0, False, 6, 2, clparam_num_antennas, 1, starting_channel,
num_channels, clparam_integration_frames, clparam_antenna_list,
True, output_file, 0, True, clparam_snap_sync, clparam_object_name,
clparam_starting_chan_freq, clparam_channel_width,
clparam_no_output, clparam_cpu_integration)
if clparam_enable_affinity:
# So with affinity here, we're just trying to ensure NUMA doesn't move us off
# where our memory was allocated. So we're going to try to be smart about
# allocating here. We'll set affinity to all of the cores on each processor till
# we've "recommended" a full set.
num_nodes = numa.get_max_node() + 1
# core_pairs = []
cpu_core_list = []
cores_per_cpu = 0
cores_per_cpu_2 = 0
for cur_node in range(0, num_nodes):
cpu_to_node = list(numa.node_to_cpus(cur_node))
cpu_core_list.append(cpu_to_node)
if cores_per_cpu == 0:
cores_per_cpu = len(cpu_to_node)
cores_per_cpu_2 = cores_per_cpu // 2
print("CPU" + str(cur_node) + " has " + str(len(cpu_to_node)) +
" cores: " + str(cpu_to_node))
#i = 0
#for cur_cpu in cpu_to_node:
# if i % 2 == 0:
# cpu_pair = [cur_cpu]
# else:
# cpu_pair.append(cur_cpu)
# core_pairs.append(cpu_pair)
#
# i += 1
#print("Setting xEngine affinity to cores " + str(core_pairs[0]))
#self.clenabled_clXEngine_0.set_processor_affinity(core_pairs[0])
#core_pairs = core_pairs[1:]
# or to all cores on CPU0
if num_nodes > 1:
self.clenabled_clXEngine_0.set_processor_affinity(
cpu_core_list[0])
self.antenna_list = []
for i in range(0, clparam_num_antennas):
if i == 0:
input_file = '/home/sonata/casa_pcap/snap_2_ant_1f.pcap'
input_port = clparam_base_port + i
elif i == 1:
input_file = '/home/sonata/casa_pcap/snap_7_ant_3c.pcap'
input_port = clparam_base_port + i
else:
input_file = '/home/sonata/casa_pcap/snap_8_ant_4g.pcap'
input_port = 10002
new_ant = ata.snap_source(input_port, 1, True, False, False,
starting_channel, ending_channel, 3,
input_file, False, True, '224.1.1.10',
False)
if clparam_enable_affinity and num_nodes > 1:
cpu2 = cores_per_cpu_2 - 2 + cores_per_cpu_2
if i < cores_per_cpu_2 - 2:
# Subtract 2 for the xengine on the first node.
print("Setting affinity for PCAP " + input_file +
" to CPU0")
new_ant.set_processor_affinity(cpu_core_list[0])
elif i < cpu2:
print("Setting affinity for PCAP " + input_file +
" to CPU1")
new_ant.set_processor_affinity(cpu_core_list[1])
else:
# just balance
index = i % 2
print("Setting affinity for PCAP " + input_file +
" to CPU" + str(index))
new_ant.set_processor_affinity(cpu_core_list[index])
##################################################
# Connections
##################################################
self.msg_connect((self.clenabled_clXEngine_0, 'sync'),
(new_ant, 'sync'))
self.connect((new_ant, 0), (self.clenabled_clXEngine_0, i))
self.antenna_list.append(new_ant)
def get_dst_numa_node_from_pcpu(self, pcpu_id):
#module numa has not implemented numa_node_of_cpu() call of numa(3) library
for i in range(0, numa.get_max_node() + 1):
if pcpu_id in numa.node_to_cpus(i):
return i
def onlinecpu():
import numa
for node in range(0, numa.get_max_node() + 1):
for cpu in sorted(numa.node_to_cpus(node)):
yield str(cpu)
def test_node_size(self):
for node in range(numa.get_max_node()+1):
print 'Node: %d, size: %r' % (node, numa.get_node_size(node))
"figureOpen": 0,
"figureName": "libfigureSL240.so",
"figureParams": None,
"reconName": "libreconmvm.so",
"fluxThreshold": 0,
"printUnused": 1,
"useBrightest": 0,
"figureGain": 1,
"decayFactor": None, #used in libreconmvm.so
"reconlibOpen": 1,
"maxAdapOffset": 10,
"noPrePostThread": 0,
}
nthr = control.get("ncamThreads").sum()
nnodes = numa.get_max_node() + 1
threadToNuma = numpy.zeros(nthr, numpy.int32)
sf = control.get("subapFlag")
nsub = sf.size
sa = numpy.zeros(nsub, numpy.int32)
#divide all threads equally....
sa[:] = nthr - 1
start = 0
for i in range(nnodes):
control["numa%d" % i] = {}
for i in range(nthr):
end = start + nsub / nthr
sa[start:end] = i
start = end
control["threadToNuma"] = threadToNuma
gainReconmxT = control["rmx"].transpose().astype(numpy.float32)
def test_node_size(self):
for node in range(numa.get_max_node()+1):
print('Node: %d, size: %r' % (node, numa.get_node_size(node)))
def onlinecpu():
import numa
for node in range(0,numa.get_max_node()+1):
for cpu in sorted(numa.node_to_cpus(node)):
yield str(cpu)
def __init__(self, **kwargs):
test_module = kwargs.pop('test_module')
viewers = kwargs.pop('viewers', False)
n_nodes = kwargs.pop('n_nodes')
n_workers = kwargs.pop('n_workers')
partition = kwargs.pop('partition')
do_numa = kwargs.pop('numa')
log_file = kwargs.pop('log_file')
input_files = kwargs.pop('input_files')
self.__logger = kwargs.pop('logger')
from collections import defaultdict
self.__results = defaultdict(list)
import numa
max_nodes = numa.get_max_node() + 1
if do_numa:
if n_nodes > max_nodes:
print(
'WARNING: %d NUMA nodes have been requested to be used, ' +
'but only %d are available and will be used.')
n_nodes = max_nodes
n_nodes = max_nodes if n_nodes == -1 else n_nodes
CalibLauncher = NumaLauncher
else:
n_nodes = 1 if n_nodes == -1 else n_nodes
CalibLauncher = Launcher
if n_workers != -1:
n_work = n_workers
else:
import multiprocessing
n_work = int(
float(multiprocessing.cpu_count()) / float(max_nodes) + 0.5)
tms = __import__("CalibTests.TestModules", globals(), locals(),
[test_module])
tm = getattr(tms, test_module)
import os
directory = os.path.dirname(input_files[0])
self.__logger.info("Will start %d workers on %d nodes." %
(n_work, n_nodes))
launcher = Launcher(partition, self.__logger)
self.__components = [
(LogServer, dict(launcher=launcher, priority=1)),
(Iterator,
dict(launcher=launcher, priority=3,
python=tm.Iterator(directory))),
(Analyser,
dict(n_nodes=n_nodes,
n_workers=n_work,
python=tm.Analyzer(input_files),
launcher=CalibLauncher(partition, self.__logger),
priority=4))
]
if log_file:
self.__components.append((LogViewer,
dict(launcher=launcher,
priority=2,
output_file=log_file,
timing_width=90,
split_regex='CALIBWORK_(\d+)')))
if viewers:
self.__components.append(
(LogViewer, dict(launcher=launcher, priority=2)))
self.__calib_process = CalibrationSteering("CalibrationTest",
self.__logger,
self.__components)
# To get an idea of overall CPU usage
from OfflineOnline.BenchmarkProcess import CPUMeasurer
self.__cpu_process = CPUMeasurer(self.__logger)
