Skip to content

Commit

Permalink
Merge pull request #159 from weka/vince/res_gen_protocols_mem
Browse files Browse the repository at this point in the history 
Vince/res gen protocols mem
  • Loading branch information
@vince-weka
vince-weka authored Jun 16, 2023
2 parents 87cea31 + 3c89449 commit 803171f
Show file tree
Hide file tree
Showing 2 changed files with 30 additions and 20 deletions.
50 changes: 30 additions & 20 deletions install/resources_generator.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -440,6 +440,9 @@ def add_arguments(self, actions):
parser.add_argument("--spare-memory", default=0, type=_validate_memory_size,
help="Specify how much memory should be reserved for non-weka requirements, "
"argument should be value and unit without whitespace (i.e 10GiB, 1024B, 5TiB etc.)")
parser.add_argument("--protocols-memory", default=0, type=_validate_memory_size,
help="Specify how much memory should be reserved for Protocols requirements, "
"argument should be value and unit without whitespace (i.e 10GiB, 1024B, 5TiB etc.)")
parser.add_argument("--compute-memory", default=0, type=_validate_memory_size,
help="Specify how much total memory should be allocated for COMPUTE, "
"argument should be value and unit without whitespace (i.e 10GiB, 1024B, 5TiB etc.)")
Expand Down Expand Up @@ -803,32 +806,39 @@ def _estimate_nodes_resident_memory_size(self, nodes=None):
mgmt_count = len(self.num_containers_by_role)
logger.debug("_estimate_nodes_resident_memory_size: compute_count=%s, drive_count=%s, frontend=%s, mgmt_count=%s",
compute_count, drive_count, frontend_count, mgmt_count)
return compute_count * WEKANODE_BUCKET_PROCESS_MEMORY + drive_count * WEKANODE_SSD_PROCESS_MEMORY +\
frontend_count * WEKANODE_FRONTEND_PROCESS_MEMORY + mgmt_count * WEKANODE_MANAGER_PROCESS_MEMORY / len(self.numa_nodes_info)
return (compute_count * WEKANODE_BUCKET_PROCESS_MEMORY) + (drive_count * WEKANODE_SSD_PROCESS_MEMORY) + \
(frontend_count * WEKANODE_FRONTEND_PROCESS_MEMORY) + \
(mgmt_count * WEKANODE_MANAGER_PROCESS_MEMORY / len(self.numa_nodes_info))

def _get_reserved_memory(self):
"""Return how many bytes cannot be allocated for wekanodes' hugepages"""
protocols_reserved_memory = 8 * GiB
#protocols_reserved_memory = 8 * GiB # vince - This should be a parameter? (it's too small)
RDMA_reserved_memory = 2 * GiB
max_reserved_portion_denom = 5 # 20% of total memory
# max_reserved_portion_denom = 5 # 20% of total memory (huh?)
total_memory_bytes = self._get_total_memory_bytes()
logger.debug("Total memory: %s MiB", total_memory_bytes / MiB)
os_reserved_memory = auto_os_reserved_memory = self._get_os_reserved_memory(total_memory_bytes)
# total_reserve = auto_os_reserved_memory = self._get_os_reserved_memory(total_memory_bytes)
total_reserve = auto_os_reserved_memory = MIN_OS_RESERVED_MEMORY
if self.args.spare_memory:
if self.args.spare_memory < auto_os_reserved_memory:
logger.warning("The spare-memory requested: %s bytes is lower than the minimum recommended memory for this "
"machine: %s bytes", self.args.spare_memory, auto_os_reserved_memory)
self.check_if_should_continue()
os_reserved_memory = self.args.spare_memory
total_reserve = os_reserved_memory + protocols_reserved_memory
total_reserve += self.args.spare_memory
if self.args.protocols_memory:
total_reserve += self.args.protocols_memory
logger.debug(f'_get_reserved_memory: protocols memory {round(self.args.protocols_memory / GiB, 1)} GiB')

#total_reserve = os_reserved_memory
if not self.args.no_rdma:
total_reserve += RDMA_reserved_memory
logger.debug("_get_reserved_memory: reserving %s MiB for RDMA", RDMA_reserved_memory / MiB)
if not self.args.spare_memory and total_reserve > total_memory_bytes / max_reserved_portion_denom:
fixed_total_reserve = total_memory_bytes / max_reserved_portion_denom
logger.warning("we need reserves of %s GiB which is too much (host has %s GiB), we'll settle with %s GiB reserved"
, total_reserve / GiB, total_memory_bytes / GiB, fixed_total_reserve / GiB)
total_reserve = fixed_total_reserve
#if not self.args.spare_memory and total_reserve > total_memory_bytes / max_reserved_portion_denom:
# fixed_total_reserve = total_memory_bytes / max_reserved_portion_denom # vince- WRONG! High-mem systems this is ok
# logger.warning("we need reserves of %s GiB which is too much (host has %s GiB), we'll settle with %s GiB reserved"
# , total_reserve / GiB, total_memory_bytes / GiB, fixed_total_reserve / GiB)
# total_reserve = fixed_total_reserve
logger.debug(f'_get_reserved_memory: total reserved memory {round(total_reserve / GiB, 1)} GiB')
return total_reserve

def _get_hugepages_memory_per_compute_node(self, numa_total_memory, io_nodes):
Expand All @@ -839,7 +849,7 @@ def _get_hugepages_memory_per_compute_node(self, numa_total_memory, io_nodes):
logger.debug("_get_hugepages_memory_per_compute_node no compute nodes, will return 0")
return 0
wekanodes_memory = self._estimate_nodes_resident_memory_size(io_nodes) * wekanodes_memory_factor
logger.debug("_get_hugepages_memory_per_compute_node WEKANODES MEMORY: %s GiB", wekanodes_memory / GiB)
logger.debug("_get_hugepages_memory_per_compute_node WEKANODES RSS MEMORY: %s GiB", round(wekanodes_memory / GiB, 2))
available = numa_total_memory - wekanodes_memory
if available <= 0:
logger.warning("_get_hugepages_memory_per_compute_node not enough available memory, will set memory to the minimal")
Expand Down Expand Up @@ -927,22 +937,22 @@ def set_memory(self):
if self.args.compute_memory:
logger.error("minimal-memory and compute-memory cannot be specified together")
quit(1)
compute_node_hugaepages_memory = DEFAULT_NODE_HUGEPAGES_MEMORY_BYTES
compute_node_hugepages_memory = DEFAULT_NODE_HUGEPAGES_MEMORY_BYTES
elif self.args.weka_hugepages_memory:
compute_memory = self._get_compute_mem_from_specified_total()
compute_node_hugaepages_memory = self._get_validated_compute_memory_arg(compute_memory)
compute_node_hugepages_memory = self._get_validated_compute_memory_arg(compute_memory)
else:
compute_node_hugaepages_memory = self._get_compute_slot_memory_requirement()
compute_node_hugepages_memory = self._get_compute_slot_memory_requirement()
if self.args.compute_memory: # user specified compute-memory
compute_node_hugaepages_memory = self._get_validated_compute_memory_arg(self.args.compute_memory)
compute_node_hugepages_memory = self._get_validated_compute_memory_arg(self.args.compute_memory)

for role in self.containers:
for container in self.containers[role]:
if role == COMPUTE_ROLE:
compute_nodes_cound = len(list(filter(lambda n: n.is_compute(), container.nodes.values())))
memory = compute_node_hugaepages_memory * compute_nodes_cound
compute_nodes_count = len(list(filter(lambda n: n.is_compute(), container.nodes.values())))
memory = compute_node_hugepages_memory * compute_nodes_count
container.memory = memory
logger.info("allocating %s GiB for %s container, (%s nodes)", memory / GiB, role, compute_nodes_cound)
logger.info("allocating %s GiB for %s container, (%s nodes)", memory / GiB, role, compute_nodes_count)
else:
container.memory = 0

Expand Down
Binary file modified install/wekaconfig
View file
Binary file not shown.

0 comments on commit 803171f

Please sign in to comment.