3 from __future__ import absolute_import, print_function
8 from . import clientactor
9 from .config import ARVADOS_ERRORS
11 class ServerCalculator(object):
12 """Generate cloud server wishlists from an Arvados job queue.
14 Instantiate this class with a list of cloud node sizes you're willing to
15 use, plus keyword overrides from the configuration. Then you can pass
16 job queues to servers_for_queue. It will return a list of node sizes
17 that would best satisfy the jobs, choosing the cheapest size that
18 satisfies each job, and ignoring jobs that can't be satisfied.
21 class CloudSizeWrapper(object):
22 def __init__(self, real_size, node_mem_scaling, **kwargs):
24 for name in ['id', 'name', 'ram', 'disk', 'bandwidth', 'price',
26 setattr(self, name, getattr(self.real, name))
27 self.cores = kwargs.pop('cores')
28 # libcloud disk sizes are in GB, Arvados/SLURM are in MB
29 # multiply by 1000 instead of 1024 to err on low side
30 self.scratch = self.disk * 1000
31 self.ram = int(self.ram * node_mem_scaling)
32 for name, override in kwargs.iteritems():
33 if not hasattr(self, name):
34 raise ValueError("unrecognized size field '%s'" % (name,))
35 setattr(self, name, override)
37 if self.price is None:
38 raise ValueError("Required field 'price' is None")
40 def meets_constraints(self, **kwargs):
41 for name, want_value in kwargs.iteritems():
42 have_value = getattr(self, name)
43 if (have_value != 0) and (have_value < want_value):
48 def __init__(self, server_list, max_nodes=None, max_price=None,
49 node_mem_scaling=0.95):
50 self.cloud_sizes = [self.CloudSizeWrapper(s, node_mem_scaling, **kws)
51 for s, kws in server_list]
52 self.cloud_sizes.sort(key=lambda s: s.price)
53 self.max_nodes = max_nodes or float('inf')
54 self.max_price = max_price or float('inf')
55 self.logger = logging.getLogger('arvnodeman.jobqueue')
56 self.logged_jobs = set()
58 self.logger.info("Using cloud node sizes:")
59 for s in self.cloud_sizes:
60 self.logger.info(str(s.__dict__))
63 def coerce_int(x, fallback):
66 except (TypeError, ValueError):
69 def cloud_size_for_constraints(self, constraints):
70 want_value = lambda key: self.coerce_int(constraints.get(key), 0)
71 wants = {'cores': want_value('min_cores_per_node'),
72 'ram': want_value('min_ram_mb_per_node'),
73 'scratch': want_value('min_scratch_mb_per_node')}
74 for size in self.cloud_sizes:
75 if size.meets_constraints(**wants):
79 def servers_for_queue(self, queue):
83 seen_jobs.add(job['uuid'])
84 constraints = job['runtime_constraints']
85 want_count = max(1, self.coerce_int(constraints.get('min_nodes'), 1))
86 cloud_size = self.cloud_size_for_constraints(constraints)
87 if cloud_size is None:
88 if job['uuid'] not in self.logged_jobs:
89 self.logged_jobs.add(job['uuid'])
90 self.logger.debug("job %s not satisfiable", job['uuid'])
91 elif (want_count <= self.max_nodes) and (want_count*cloud_size.price <= self.max_price):
92 servers.extend([cloud_size.real] * want_count)
93 self.logged_jobs.intersection_update(seen_jobs)
96 def cheapest_size(self):
97 return self.cloud_sizes[0]
99 def find_size(self, sizeid):
100 for s in self.cloud_sizes:
105 class JobQueueMonitorActor(clientactor.RemotePollLoopActor):
106 """Actor to generate server wishlists from the job queue.
108 This actor regularly polls Arvados' job queue, and uses the provided
109 ServerCalculator to turn that into a list of requested node sizes. That
110 list is sent to subscribers on every poll.
113 CLIENT_ERRORS = ARVADOS_ERRORS
115 def __init__(self, client, timer_actor, server_calc,
116 jobs_queue, slurm_queue, *args, **kwargs):
117 super(JobQueueMonitorActor, self).__init__(
118 client, timer_actor, *args, **kwargs)
119 self.jobs_queue = jobs_queue
120 self.slurm_queue = slurm_queue
121 self._calculator = server_calc
128 elif u in ("G", "g"):
129 return float(v) * 2**10
130 elif u in ("T", "t"):
131 return float(v) * 2**20
132 elif u in ("P", "p"):
133 return float(v) * 2**30
137 def _send_request(self):
140 # cpus, memory, tempory disk space, reason, job name
141 squeue_out = subprocess.check_output(["squeue", "--state=PENDING", "--noheader", "--format=%c|%m|%d|%r|%j"])
142 for out in squeue_out.splitlines():
144 cpu, ram, disk, reason, jobname = out.split("|", 4)
145 if ("ReqNodeNotAvail" in reason) or ("Resources" in reason):
148 "runtime_constraints": {
149 "min_cores_per_node": cpu,
150 "min_ram_mb_per_node": self.coerce_to_mb(ram),
151 "min_scratch_mb_per_node": self.coerce_to_mb(disk)
158 queuelist.extend(self._client.jobs().queue().execute()['items'])
162 def _got_response(self, queue):
163 server_list = self._calculator.servers_for_queue(queue)
164 self._logger.debug("Calculated wishlist: %s",
165 ', '.join(s.name for s in server_list) or "(empty)")
166 return super(JobQueueMonitorActor, self)._got_response(server_list)