[HiCache] Manage buffer-mode backups per storage pool (#42633)
Co-authored-by: merrymercy <merrymercy@users.noreply.github.com> Co-authored-by: xiezhq-hermann <xiezhq-hermann@users.noreply.github.com>
This commit is contained in:
co-authored by
merrymercy
xiezhq-hermann
parent
fd62ee84f7
commit
08aad0c196
@@ -1295,15 +1295,9 @@ class HiCacheController:
|
||||
)
|
||||
hash_value, storage_hit_count = [], 0
|
||||
|
||||
storage_hit_count_tensor = torch.tensor(
|
||||
storage_hit_count, dtype=torch.int
|
||||
storage_hit_count = self._sync_prefetch_hit_query(
|
||||
operation, storage_hit_count
|
||||
)
|
||||
self._all_reduce(
|
||||
storage_hit_count_tensor,
|
||||
torch.distributed.ReduceOp.MIN,
|
||||
self.prefetch_hits_sync_groups,
|
||||
)
|
||||
storage_hit_count = storage_hit_count_tensor.item()
|
||||
|
||||
# Record the TP-synced hit count; the scheduler thread decides
|
||||
# at drain time whether to revoke (below threshold) or allocate.
|
||||
@@ -1316,6 +1310,17 @@ class HiCacheController:
|
||||
except Empty:
|
||||
continue
|
||||
|
||||
def _sync_prefetch_hit_query(self, operation, storage_hit_count: int) -> int:
|
||||
"""Rank-reduce a hit query (MIN) so every rank agrees on the usable
|
||||
prefix. Runs for every operation, hit or miss."""
|
||||
hit_count = torch.tensor(storage_hit_count, dtype=torch.int)
|
||||
self._all_reduce(
|
||||
hit_count,
|
||||
torch.distributed.ReduceOp.MIN,
|
||||
self.prefetch_hits_sync_groups,
|
||||
)
|
||||
return int(hit_count.item())
|
||||
|
||||
def write_storage(
|
||||
self,
|
||||
host_indices: torch.Tensor,
|
||||
|
||||
Reference in New Issue
Block a user