Rss ladder py (#118)

Added: rss-ladder now supports --offset and --cpu options.

If you have 2 NICs with 4 queues and 1 socket with 8 cpus, you may be want
distribution like this:

eth0: [0, 1, 2, 3]
eth1: [4, 5, 6, 7]

so run:

```
rss-ladder-test eth0
rss-ladder-test --offset=4 eth1
```

If you have bugged NIC missing packets while RSS used on multiple CPUs
(it may be caused by DMA/RAM problems) you may use the following workaround:

bind all the NIC queues to **one** cpu, using:

```
rss-ladder --cpus 0 eth0
```

and distribute processing workload among other available CPUs using RPS.

netutils_linux_tuning/rss_ladder.py

@@ -6,15 +6,20 @@
 from os.path import join
 from argparse import ArgumentParser
 from six import iteritems, print_
+from six.moves import xrange
 from netutils_linux_monitoring.numa import Numa
 from netutils_linux_hardware.assessor_math import any2int
 
+MAX_QUEUE_PER_DEVICE = 16
+
 
 class RSSLadder(object):
+ """ Distributor of queues' interrupts by multiple CPUs """
 
  def __init__(self):
  interrupts_file = '/proc/interrupts'
  self.options = self.parse_options()
+ print_(self.options)
  lscpu_output = None
  if self.options.test_dir:
  interrupts_file = join(self.options.test_dir, "interrupts")
@@ -26,7 +31,8 @@ def __init__(self):
  if isinstance(lscpu_output, bytes):
  lscpu_output = str(lscpu_output)
  self.interrupts = open(interrupts_file).readlines()
- self.numa = Numa(lscpu_output=lscpu_output)
+ if not self.options.cpus: # no need to detect topology if user gave us cpu list
+ self.numa = Numa(lscpu_output=lscpu_output)
  for postfix in sorted(self.queue_postfixes_detect()):
  self.smp_affinity_list_apply(self.smp_affinity_list_make(postfix))
 
@@ -41,6 +47,8 @@ def parse_options():
  help="If you have 2 NICs with 4 queues and 1 socket with 8 cpus, you may be want "
  "distribution like this: eth0: [0, 1, 2, 3]; eth1: [4, 5, 6, 7]; "
  "so run: rss-ladder-test eth0; rss-ladder-test --offset=4 eth1")
+ parser.add_argument('-c', '--cpus', help='Explicitly define list of CPUs for binding NICs queues', type=int,
+ nargs='+')
  parser.add_argument('dev', type=str)
  parser.add_argument('socket', nargs='?', type=int, default=0)
  return parser.parse_args()
@@ -70,12 +78,20 @@ def smp_affinity_list_make(self, postfix):
  print_("- distribute interrupts of {0} ({1}) on socket {2}".format(
  self.options.dev, postfix, self.options.socket))
  queue_regex = r'{0}{1}[^ \n]+'.format(self.options.dev, postfix)
- socket_cpus = [k for k, v in iteritems(self.numa.socket_layout) if v == self.options.socket] * 4
- socket_cpus.reverse()
+ if self.options.cpus:
+ # 16 is in case of someone decide to bind up to 16 queues to one CPU for cache locality
+ # and manually distribute workload by RPS to other CPUs.
+ rss_cpus = self.options.cpus * MAX_QUEUE_PER_DEVICE
+ else:
+ cpus = [k for k, v in iteritems(self.numa.socket_layout) if v == self.options.socket]
+ rss_cpus = cpus * MAX_QUEUE_PER_DEVICE
+ rss_cpus.reverse()
+ for _ in xrange(self.options.offset):
+ rss_cpus.pop()
  for line in self.interrupts:
  queue_name = re.findall(queue_regex, line)
  if queue_name:
- yield any2int(line.split()[0]), queue_name[0], socket_cpus.pop()
+ yield any2int(line.split()[0]), queue_name[0], rss_cpus.pop()
 
  def smp_affinity_list_apply(self, smp_affinity):
  """

setup.py

@@ -14,7 +14,7 @@ def read(*paths):
 
 setuptools.setup(
  name='netutils-linux',
- version='2.1.0',
+ version='2.1.1',
  author='Oleg Strizhechenko',
  author_email='[email protected]',
  license='MIT',