def next(root, jobs, max_weight=None, stale_seconds=None):
if g.rank() == 0:
j = get_next_name(root, jobs, max_weight,
stale_seconds).encode("utf-8")
else:
j = bytes()
j_name = g.broadcast(0, j).decode("utf-8")
for j in jobs:
if j.name == j_name:
g.message(f"""
--------------------------------------------------------------------------------
Start job {j.name}
--------------------------------------------------------------------------------
""")
t0 = g.time()
j(root)
t1 = g.time()
g.message(f"""
--------------------------------------------------------------------------------
Completed {j.name} in {t1-t0} seconds
--------------------------------------------------------------------------------
""")
return j
return None
def global_sum(self):
me = str([self.basis.to_array(), self[:]]).encode("utf-8")
r = tensor(basis([]), self.n_parallel)
for i in range(gpt.ranks()):
b, t = eval(gpt.broadcast(i, me).decode("utf-8"))
ti = tensor(basis(b), self.n_parallel)
ti[:] = t
r = r + ti
return r
def read_numpy(self, start, end):
if gpt.rank() == 0:
self.glb.seek(start, 0)
crc32_compare = int.from_bytes(self.glb.read(4),
byteorder='little')
data = self.glb.read(end - start - 4)
else:
data = None
crc32_compare = None
data = gpt.broadcast(0, data)
crc32_computed = gpt.crc32(memoryview(data))
if not crc32_compare is None:
assert (crc32_computed == crc32_compare)
return numpy.load(io.BytesIO(data))
return None
if g.rank() == 0:
first_job = get_job()
run_jobs = str(
list(
filter(
lambda x: x is not None,
[first_job] +
[get_job(first_job[2]) for i in range(1, jobs_per_run)],
))).encode("utf-8")
else:
run_jobs = bytes()
run_jobs = eval(g.broadcast(0, run_jobs).decode("utf-8"))
# every node now knows what to do
g.message("""
================================================================================
HVP connected run on summit ; this run will attempt:
================================================================================
""")
for group, job, conf, jid, n in run_jobs:
g.message(f"""
Job {jid} / {n} : configuration {conf}, job tag {job}
""")
# configuration needs to be the same for all jobs, so load eigenvectors and configuration
