Fix #1638: double CRC32C computation in RBD layer #1712
Conversation
- Add skip_rbd_crc_if_transport_digest config option to disable redundant RBD CRC computation - Implement config-driven logic in GatewayServer._initialize_rbd_crc32c() - Add unit tests validating all config scenarios Signed-off-by: Annmool <aydv.267@gmail.com>
Annmool
force-pushed
the
double-computation-problem
branch
from
ead43c9 to
dd5e8e7
Compare
|
|
||
| effective_rbd_crc = bool(rbd_with_crc32c) | ||
| if skip_if_transport_digest: | ||
| effective_rbd_crc = False |
There was a problem hiding this comment.
This seems to be adding a new config option (skip_rbd_crc_if_transport_digest) whose sole purpose is to override the existing config option (rbd_with_crc32c) in case it's set to true. Is there anything that is preventing the operator from manipulating rbd_with_crc32c option directly?
There was a problem hiding this comment.
I also don't understand what is the idea of this PR given that #1429 is handling the same issue exactly and is already merged.
There was a problem hiding this comment.
@idryomov You make a valid point. Looking at the current behavior, rbd_with_crc32c already defaults to False, so operators who want to skip RBD CRC computation can simply leave it unset or explicitly set it to False. The new skip_rbd_crc_if_transport_digest option is somewhat redundant from a functional perspective.
what we can do is
- Remove the new
skip_rbd_crc_if_transport_digestoption - Just rely on rbd_with_crc32c = False (already the default)
- Update documentation to clarify: operators who understand the optimization can disable RBD CRC by not enabling it
There was a problem hiding this comment.
ok so I guess it need to be closed right?
3 participants
Fix #1638
Problem: The NVMe-oF transport layer and RBD block device layer were both computing CRC32C checksums on the same data, wasting CPU cycles.
Solution
Config-Driven Optimization (server.py)
Modified
_initialize_rbd_crc32c()method to read a new config option:skip_rbd_crc_if_transport_digestWhen enabled, this flag disables RBD-side CRC32C computation entirely
The logic respects explicit
rbd_with_crc32csettings but allows the skip flag to override them for safetyAdded informative logging so operators know when the optimization is active
Configuration Option (ceph-nvmeof.conf)
Added
skip_rbd_crc_if_transport_digest = True/Falseoption under the [spdk] sectionIncluded safety warnings in comments explaining when to enable (trusted stacks only, no data transformations)
Result: Operators can now eliminate redundant CRC32C recomputation by setting one config flag, reducing CPU overhead on write-heavy workloads while maintaining safety through explicit opt-in and clear warnings.