DB/WALL and RGW index on the same NVME
Hi! I'm working on a POC cluster setup dedicated to backup app writing objects via s3 (large objects, up to 1TB transferred via multipart upload process). Initial setup is 18 storage nodes (12HDDs + 1 NVME card for DB/WALL) + EC pool. Plan is to use cephadm. I'd like to follow good practice and put the RGW index pool on a no-rotation drive. Question is how to do it? - replace a few HDDs (1 per node) with a SSD (how many? 4-6-8?) - reserve space on NVME drive on each node, create lv based OSD and let rgb index use the same NVME drive as DB/WALL Thoughts? -- Lukasz
Hi Lukasz, RGW uses Omap objects for the index pool; Omaps are stored in Rocksdb database of each osd, not on the actual index pool, so by putting DB/WALL on an NVMe as you mentioned, you are already configuring the index pool on a non-rotational drive, you don't need to do anything else. You just need to size your DB/WALL partition accordingly. For RGW/object storage, a good starting point for the DB/Wall sizing is 4%. Example of Omap entries in the index pool using 0 bytes, as they are stored in Rocksdb: # rados -p default.rgw.buckets.index listomapkeys .dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 file1 file2 file4 file10 rados df -p default.rgw.buckets.index POOL_NAME USED OBJECTS CLONES COPIES MISSING_ON_PRIMARY UNFOUND DEGRADED RD_OPS RD WR_OPS WR USED COMPR UNDER COMPR default.rgw.buckets.index 0 B 11 0 33 0 0 0 208 207 KiB 41 20 KiB 0 B 0 B # rados -p default.rgw.buckets.index stat .dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 default.rgw.buckets.index/.dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 mtime 2022-12-20T07:32:11.000000-0500, size 0 On Sun, Apr 7, 2024 at 10:06 PM Lukasz Borek <lukasz@borek.org.pl> wrote:
Hi!
I'm working on a POC cluster setup dedicated to backup app writing objects via s3 (large objects, up to 1TB transferred via multipart upload process).
Initial setup is 18 storage nodes (12HDDs + 1 NVME card for DB/WALL) + EC pool. Plan is to use cephadm.
I'd like to follow good practice and put the RGW index pool on a no-rotation drive. Question is how to do it?
- replace a few HDDs (1 per node) with a SSD (how many? 4-6-8?) - reserve space on NVME drive on each node, create lv based OSD and let rgb index use the same NVME drive as DB/WALL
Thoughts?
-- Lukasz _______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io
Thanks for clarifying. So redhat doc <https://access.redhat.com/documentation/en-us/red_hat_ceph_storage/2/html-single/ceph_object_gateway_for_production/index#adv-rgw-hardware-bucket-index> is outdated? 3.6. Selecting SSDs for Bucket Indexes When selecting OSD hardware for use with a Ceph Object Gateway—irrespective
of the use case—Red Hat recommends considering an OSD node that has at least one SSD drive used exclusively for the bucket index pool. This is particularly important when buckets will contain a large number of objects.
A bucket index entry is approximately 200 bytes of data, stored as an
object map (omap) in leveldb. While this is a trivial amount of data, some uses of Ceph Object Gateway can result in tens or hundreds of millions of objects in a single bucket. By mapping the bucket index pool to a CRUSH hierarchy of SSD nodes, the reduced latency provides a dramatic performance improvement when buckets contain very large numbers of objects.
Important In a production cluster, a typical OSD node will have at least one SSD for the bucket index, AND at least on SSD for the journal.
Current utilisation is what osd df command shows in OMAP field?: root@cephbackup:/# ceph osd df
ID CLASS WEIGHT REWEIGHT SIZE RAW USE DATA OMAP META AVAIL %USE VAR PGS STATUS 0 hdd 7.39870 1.00000 7.4 TiB 894 GiB 769 GiB 1.5 MiB 3.4 GiB 6.5 TiB 11.80 1.45 40 up 1 hdd 7.39870 1.00000 7.4 TiB 703 GiB 578 GiB 6.0 MiB 2.9 GiB 6.7 TiB 9.27 1.14 37 up 2 hdd 7.39870 1.00000 7.4 TiB 700 GiB 576 GiB 3.1 MiB 3.1 GiB 6.7 TiB 9.24 1.13 39 up
On Mon, 8 Apr 2024 at 08:42, Daniel Parkes <dparkes@redhat.com> wrote:
Hi Lukasz,
RGW uses Omap objects for the index pool; Omaps are stored in Rocksdb database of each osd, not on the actual index pool, so by putting DB/WALL on an NVMe as you mentioned, you are already configuring the index pool on a non-rotational drive, you don't need to do anything else.
You just need to size your DB/WALL partition accordingly. For RGW/object storage, a good starting point for the DB/Wall sizing is 4%.
Example of Omap entries in the index pool using 0 bytes, as they are stored in Rocksdb:
# rados -p default.rgw.buckets.index listomapkeys .dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 file1 file2 file4 file10
rados df -p default.rgw.buckets.index POOL_NAME USED OBJECTS CLONES COPIES MISSING_ON_PRIMARY UNFOUND DEGRADED RD_OPS RD WR_OPS WR USED COMPR UNDER COMPR default.rgw.buckets.index 0 B 11 0 33 0 0 0 208 207 KiB 41 20 KiB 0 B 0 B
# rados -p default.rgw.buckets.index stat .dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 default.rgw.buckets.index/.dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 mtime 2022-12-20T07:32:11.000000-0500, size 0
On Sun, Apr 7, 2024 at 10:06 PM Lukasz Borek <lukasz@borek.org.pl> wrote:
Hi!
I'm working on a POC cluster setup dedicated to backup app writing objects via s3 (large objects, up to 1TB transferred via multipart upload process).
Initial setup is 18 storage nodes (12HDDs + 1 NVME card for DB/WALL) + EC pool. Plan is to use cephadm.
I'd like to follow good practice and put the RGW index pool on a no-rotation drive. Question is how to do it?
- replace a few HDDs (1 per node) with a SSD (how many? 4-6-8?) - reserve space on NVME drive on each node, create lv based OSD and let rgb index use the same NVME drive as DB/WALL
Thoughts?
-- Lukasz _______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io
-- Łukasz Borek lukasz@borek.org.pl
Hi, Yes, that documentation you are linking is from Ceph 3.x with Filestore, With Bluestore this is no longer the case, the link to the latest Red Hat doc version is here: https://access.redhat.com/documentation/en-us/red_hat_ceph_storage/7/html-si... I see they have this block of text there: "For Red Hat Ceph Storage running Bluestore, Red Hat recommends deploying an NVMe drive as a block.db device, rather than as a separate pool. Ceph Object Gateway index data is written only into an object map (OMAP). OMAP data for BlueStore resides on the block.db device on an OSD. When an NVMe drive functions as a block.db device for an HDD OSD and when the index pool is backed by HDD OSDs, the index data will ONLY be written to the block.db device. As long as the block.db partition/lvm is sized properly at 4% of block, this configuration is all that is needed for BlueStore." On Mon, Apr 8, 2024 at 12:02 PM Lukasz Borek <lukasz@borek.org.pl> wrote:
Thanks for clarifying.
So redhat doc <https://access.redhat.com/documentation/en-us/red_hat_ceph_storage/2/html-single/ceph_object_gateway_for_production/index#adv-rgw-hardware-bucket-index> is outdated?
3.6. Selecting SSDs for Bucket Indexes
When selecting OSD hardware for use with a Ceph Object
Gateway—irrespective of the use case—Red Hat recommends considering an OSD node that has at least one SSD drive used exclusively for the bucket index pool. This is particularly important when buckets will contain a large number of objects.
A bucket index entry is approximately 200 bytes of data, stored as an
object map (omap) in leveldb. While this is a trivial amount of data, some uses of Ceph Object Gateway can result in tens or hundreds of millions of objects in a single bucket. By mapping the bucket index pool to a CRUSH hierarchy of SSD nodes, the reduced latency provides a dramatic performance improvement when buckets contain very large numbers of objects.
Important In a production cluster, a typical OSD node will have at least one SSD for the bucket index, AND at least on SSD for the journal.
Current utilisation is what osd df command shows in OMAP field?:
root@cephbackup:/# ceph osd df
ID CLASS WEIGHT REWEIGHT SIZE RAW USE DATA OMAP META AVAIL %USE VAR PGS STATUS 0 hdd 7.39870 1.00000 7.4 TiB 894 GiB 769 GiB 1.5 MiB 3.4 GiB 6.5 TiB 11.80 1.45 40 up 1 hdd 7.39870 1.00000 7.4 TiB 703 GiB 578 GiB 6.0 MiB 2.9 GiB 6.7 TiB 9.27 1.14 37 up 2 hdd 7.39870 1.00000 7.4 TiB 700 GiB 576 GiB 3.1 MiB 3.1 GiB 6.7 TiB 9.24 1.13 39 up
On Mon, 8 Apr 2024 at 08:42, Daniel Parkes <dparkes@redhat.com> wrote:
Hi Lukasz,
RGW uses Omap objects for the index pool; Omaps are stored in Rocksdb database of each osd, not on the actual index pool, so by putting DB/WALL on an NVMe as you mentioned, you are already configuring the index pool on a non-rotational drive, you don't need to do anything else.
You just need to size your DB/WALL partition accordingly. For RGW/object storage, a good starting point for the DB/Wall sizing is 4%.
Example of Omap entries in the index pool using 0 bytes, as they are stored in Rocksdb:
# rados -p default.rgw.buckets.index listomapkeys .dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 file1 file2 file4 file10
rados df -p default.rgw.buckets.index POOL_NAME USED OBJECTS CLONES COPIES MISSING_ON_PRIMARY UNFOUND DEGRADED RD_OPS RD WR_OPS WR USED COMPR UNDER COMPR default.rgw.buckets.index 0 B 11 0 33 0 0 0 208 207 KiB 41 20 KiB 0 B 0 B
# rados -p default.rgw.buckets.index stat .dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 default.rgw.buckets.index/.dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 mtime 2022-12-20T07:32:11.000000-0500, size 0
On Sun, Apr 7, 2024 at 10:06 PM Lukasz Borek <lukasz@borek.org.pl> wrote:
Hi!
I'm working on a POC cluster setup dedicated to backup app writing objects via s3 (large objects, up to 1TB transferred via multipart upload process).
Initial setup is 18 storage nodes (12HDDs + 1 NVME card for DB/WALL) + EC pool. Plan is to use cephadm.
I'd like to follow good practice and put the RGW index pool on a no-rotation drive. Question is how to do it?
- replace a few HDDs (1 per node) with a SSD (how many? 4-6-8?) - reserve space on NVME drive on each node, create lv based OSD and let rgb index use the same NVME drive as DB/WALL
Thoughts?
-- Lukasz _______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io
-- Łukasz Borek lukasz@borek.org.pl
I would suggest that you might consider EC vs. replication for index data, and the latency implications. There's more than just the nvme vs. rotational discussion to entertain, especially if using the more widely spread EC modes like 8+3. It would be worth testing for your particular workload. Also make sure to factor in storage utilization if you expect to see versioning/object lock in use. This can be the source of a significant amount of additional consumption that isn't planned for initially. On Mon, Apr 8, 2024, at 01:42, Daniel Parkes wrote:
Hi Lukasz,
RGW uses Omap objects for the index pool; Omaps are stored in Rocksdb database of each osd, not on the actual index pool, so by putting DB/WALL on an NVMe as you mentioned, you are already configuring the index pool on a non-rotational drive, you don't need to do anything else.
You just need to size your DB/WALL partition accordingly. For RGW/object storage, a good starting point for the DB/Wall sizing is 4%.
Example of Omap entries in the index pool using 0 bytes, as they are stored in Rocksdb:
# rados -p default.rgw.buckets.index listomapkeys .dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 file1 file2 file4 file10
rados df -p default.rgw.buckets.index POOL_NAME USED OBJECTS CLONES COPIES MISSING_ON_PRIMARY UNFOUND DEGRADED RD_OPS RD WR_OPS WR USED COMPR UNDER COMPR default.rgw.buckets.index 0 B 11 0 33 0 0 0 208 207 KiB 41 20 KiB 0 B 0 B
# rados -p default.rgw.buckets.index stat .dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 default.rgw.buckets.index/.dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 mtime 2022-12-20T07:32:11.000000-0500, size 0
On Sun, Apr 7, 2024 at 10:06 PM Lukasz Borek <lukasz@borek.org.pl> wrote:
Hi!
I'm working on a POC cluster setup dedicated to backup app writing objects via s3 (large objects, up to 1TB transferred via multipart upload process).
Initial setup is 18 storage nodes (12HDDs + 1 NVME card for DB/WALL) + EC pool. Plan is to use cephadm.
I'd like to follow good practice and put the RGW index pool on a no-rotation drive. Question is how to do it?
- replace a few HDDs (1 per node) with a SSD (how many? 4-6-8?) - reserve space on NVME drive on each node, create lv based OSD and let rgb index use the same NVME drive as DB/WALL
Thoughts?
-- Lukasz _______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io
_______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io
My understanding is that omap and EC are incompatible, though.
On Apr 8, 2024, at 09:46, David Orman <ormandj@corenode.com> wrote:
I would suggest that you might consider EC vs. replication for index data, and the latency implications. There's more than just the nvme vs. rotational discussion to entertain, especially if using the more widely spread EC modes like 8+3. It would be worth testing for your particular workload.
Also make sure to factor in storage utilization if you expect to see versioning/object lock in use. This can be the source of a significant amount of additional consumption that isn't planned for initially.
On Mon, Apr 8, 2024, at 01:42, Daniel Parkes wrote:
Hi Lukasz,
RGW uses Omap objects for the index pool; Omaps are stored in Rocksdb database of each osd, not on the actual index pool, so by putting DB/WALL on an NVMe as you mentioned, you are already configuring the index pool on a non-rotational drive, you don't need to do anything else.
You just need to size your DB/WALL partition accordingly. For RGW/object storage, a good starting point for the DB/Wall sizing is 4%.
Example of Omap entries in the index pool using 0 bytes, as they are stored in Rocksdb:
# rados -p default.rgw.buckets.index listomapkeys .dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 file1 file2 file4 file10
rados df -p default.rgw.buckets.index POOL_NAME USED OBJECTS CLONES COPIES MISSING_ON_PRIMARY UNFOUND DEGRADED RD_OPS RD WR_OPS WR USED COMPR UNDER COMPR default.rgw.buckets.index 0 B 11 0 33 0 0 0 208 207 KiB 41 20 KiB 0 B 0 B
# rados -p default.rgw.buckets.index stat .dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 default.rgw.buckets.index/.dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 mtime 2022-12-20T07:32:11.000000-0500, size 0
On Sun, Apr 7, 2024 at 10:06 PM Lukasz Borek <lukasz@borek.org.pl> wrote:
Hi!
I'm working on a POC cluster setup dedicated to backup app writing objects via s3 (large objects, up to 1TB transferred via multipart upload process).
Initial setup is 18 storage nodes (12HDDs + 1 NVME card for DB/WALL) + EC pool. Plan is to use cephadm.
I'd like to follow good practice and put the RGW index pool on a no-rotation drive. Question is how to do it?
- replace a few HDDs (1 per node) with a SSD (how many? 4-6-8?) - reserve space on NVME drive on each node, create lv based OSD and let rgb index use the same NVME drive as DB/WALL
Thoughts?
-- Lukasz _______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io
_______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io
ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io
My understanding is that omap and EC are incompatible, though.
Reason why multipart upload is using a non-EC pool to save metadata to an omap database? On Mon, 8 Apr 2024 at 20:21, Anthony D'Atri <anthony.datri@gmail.com> wrote:
My understanding is that omap and EC are incompatible, though.
On Apr 8, 2024, at 09:46, David Orman <ormandj@corenode.com> wrote:
I would suggest that you might consider EC vs. replication for index data, and the latency implications. There's more than just the nvme vs. rotational discussion to entertain, especially if using the more widely spread EC modes like 8+3. It would be worth testing for your particular workload.
Also make sure to factor in storage utilization if you expect to see versioning/object lock in use. This can be the source of a significant amount of additional consumption that isn't planned for initially.
On Mon, Apr 8, 2024, at 01:42, Daniel Parkes wrote:
Hi Lukasz,
RGW uses Omap objects for the index pool; Omaps are stored in Rocksdb database of each osd, not on the actual index pool, so by putting DB/WALL on an NVMe as you mentioned, you are already configuring the index pool on a non-rotational drive, you don't need to do anything else.
You just need to size your DB/WALL partition accordingly. For RGW/object storage, a good starting point for the DB/Wall sizing is 4%.
Example of Omap entries in the index pool using 0 bytes, as they are stored in Rocksdb:
# rados -p default.rgw.buckets.index listomapkeys .dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 file1 file2 file4 file10
rados df -p default.rgw.buckets.index POOL_NAME USED OBJECTS CLONES COPIES MISSING_ON_PRIMARY UNFOUND DEGRADED RD_OPS RD WR_OPS WR USED COMPR UNDER COMPR default.rgw.buckets.index 0 B 11 0 33 0 0 0 208 207 KiB 41 20 KiB 0 B 0 B
# rados -p default.rgw.buckets.index stat .dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2
mtime 2022-12-20T07:32:11.000000-0500, size 0
On Sun, Apr 7, 2024 at 10:06 PM Lukasz Borek <lukasz@borek.org.pl> wrote:
Hi!
I'm working on a POC cluster setup dedicated to backup app writing objects via s3 (large objects, up to 1TB transferred via multipart upload
default.rgw.buckets.index/.dir.7fb0a3df-9553-4a76-938d-d23711e67677.34162.1.2 process).
Initial setup is 18 storage nodes (12HDDs + 1 NVME card for DB/WALL) +
EC
pool. Plan is to use cephadm.
I'd like to follow good practice and put the RGW index pool on a no-rotation drive. Question is how to do it?
- replace a few HDDs (1 per node) with a SSD (how many? 4-6-8?) - reserve space on NVME drive on each node, create lv based OSD and let rgb index use the same NVME drive as DB/WALL
Thoughts?
-- Lukasz _______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io
_______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io
ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io
-- Łukasz Borek lukasz@borek.org.pl
participants (4)
-
Anthony D'Atri
-
Daniel Parkes
-
David Orman
-
Lukasz Borek