Unexpected slow read for HDD cluster (good write speed)
Hello guys! I would like to ask if somebody has already experienced a similar situation. We have a new cluster with 5 nodes with the following setup: - 128 GB of RAM - 2 cpus Intel(R) Intel Xeon Silver 4210R - 1 NVME of 2 TB for the rocks DB caching - 5 HDDs of 14TB - 1 NIC dual port of 25GiB in BOND mode. We are starting with a single dual port NIC (the bond has 50GiB in total), the design has been prepared so a new NIC can be added, and a new BOND can be created, where we intend to offload the cluster network. Therefore, logically speaking, we already configured different VLANs and networks for public and cluster traffic of Ceph. We are using Ubuntu 20.04 with Ceph Octopus. It is a standard deployment that we are used to. During our initial validations and evaluations of the cluster, we are reaching write speeds between 250-300MB/s, which would be the ballpark for this kind of setup for HDDs with the NVME as Rocks.db cache (in our experience). However, the issue is the reading process. While reading, we barely hit the mark of 100MB/s; we would expect at least something similar to the write speed. These tests are being performed in a pool with a replication factor of 3. We have already checked the disks, and they all seem to be reading just fine. The network does not seem to be the bottleneck either (checked with atop while reading/writing to the cluster). Have you guys ever encountered similar situations? Do you have any tips for us to proceed with the troubleshooting? We suspect that we are missing some small tuning detail, which is affecting the read performance only, but so far we could not pinpoint it. Any help would be much appreciated :)
While reading, we barely hit the mark of 100MB/s; we would expect at least something similar to the write speed. These tests are being performed in a pool with a replication factor of 3.
You don't even describe how you test? And why would you expect something like the write speed, as you described, that is a totally different configuration.
Den mån 20 mars 2023 kl 09:45 skrev Marc <Marc@f1-outsourcing.eu>:
While reading, we barely hit the mark of 100MB/s; we would expect at least something similar to the write speed. These tests are being performed in a pool with a replication factor of 3.
You don't even describe how you test? And why would you expect something like the write speed, as you described, that is a totally different configuration.
Yes, writes hit caches and can be async, whereas reads (at least large relevant read tests) needs to get actual data off the drives, and not from caches. Getting 100MB/s on data not in caches from hdd drives seems reasonable for a simplistic test (ie, one where you request a certain amount of data, wait for it to arrive, then read a bit more and repeat for a number of iterations). If you instead spin up 10 guests and have all of them do read tests, chances are you are going to see huge speedups. Not per-guest, but in the sum of IO that your cluster can handle, which is what ceph is aiming for. Being a good cluster storage system for cluster usage. -- May the most significant bit of your life be positive.
Yes, during my last adventure of trying to get any reasonable performance out of ceph, i realized my testing methodology was wrong. Both the kernel client and qemu have queues everywhere that make the numbers hard to understand. fio has rbd support, which gives more useful values. https://subscription.packtpub.com/book/cloud-&-networking/9781784393502/10/ch10lvl1sec112/benchmarking-ceph-rbd-using-fio frustratingly, much lower ones, showing just how slow ceph actually is. On Sat, Mar 18, 2023 at 8:59 PM Rafael Weingartner <work.ceph.user.mailing@gmail.com> wrote:
Hello guys!
I would like to ask if somebody has already experienced a similar situation. We have a new cluster with 5 nodes with the following setup:
- 128 GB of RAM - 2 cpus Intel(R) Intel Xeon Silver 4210R - 1 NVME of 2 TB for the rocks DB caching - 5 HDDs of 14TB - 1 NIC dual port of 25GiB in BOND mode.
We are starting with a single dual port NIC (the bond has 50GiB in total), the design has been prepared so a new NIC can be added, and a new BOND can be created, where we intend to offload the cluster network. Therefore, logically speaking, we already configured different VLANs and networks for public and cluster traffic of Ceph.
We are using Ubuntu 20.04 with Ceph Octopus. It is a standard deployment that we are used to. During our initial validations and evaluations of the cluster, we are reaching write speeds between 250-300MB/s, which would be the ballpark for this kind of setup for HDDs with the NVME as Rocks.db cache (in our experience). However, the issue is the reading process. While reading, we barely hit the mark of 100MB/s; we would expect at least something similar to the write speed. These tests are being performed in a pool with a replication factor of 3.
We have already checked the disks, and they all seem to be reading just fine. The network does not seem to be the bottleneck either (checked with atop while reading/writing to the cluster).
Have you guys ever encountered similar situations? Do you have any tips for us to proceed with the troubleshooting?
We suspect that we are missing some small tuning detail, which is affecting the read performance only, but so far we could not pinpoint it. Any help would be much appreciated :) _______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io
-- +4916093821054
Yes it pays off to know what to do before you do it, instead of after. If you complain about speed, is it a general unfounded complaint or did you compare ceph with similar solutions? I have really no idea what the standards are for these types of solutions. I can remember asking at such seminar that 'general' performance numbers should published for every release, so people are not having to go to the ordeal to investigate on their own. However I also get that there is a technical performance limit for distributing your data like this. I bookmarked this quite a while ago, if you are in dire need you can do some external caching for rbd's. https://docs.ceph.com/en/pacific/rbd/rbd-persistent-write-back-cache/
Yes, during my last adventure of trying to get any reasonable performance out of ceph, i realized my testing methodology was wrong. Both the kernel client and qemu have queues everywhere that make the numbers hard to understand.
fio has rbd support, which gives more useful values.
https://subscription.packtpub.com/book/cloud-&- networking/9781784393502/10/ch10lvl1sec112/benchmarking-ceph-rbd-using- fio
frustratingly, much lower ones, showing just how slow ceph actually is.
On Sat, Mar 18, 2023 at 8:59 PM Rafael Weingartner <work.ceph.user.mailing@gmail.com> wrote:
Hello guys!
I would like to ask if somebody has already experienced a similar situation. We have a new cluster with 5 nodes with the following
setup:
- 128 GB of RAM - 2 cpus Intel(R) Intel Xeon Silver 4210R - 1 NVME of 2 TB for the rocks DB caching - 5 HDDs of 14TB - 1 NIC dual port of 25GiB in BOND mode.
We are starting with a single dual port NIC (the bond has 50GiB in
the design has been prepared so a new NIC can be added, and a new BOND can be created, where we intend to offload the cluster network. Therefore, logically speaking, we already configured different VLANs and networks for public and cluster traffic of Ceph.
We are using Ubuntu 20.04 with Ceph Octopus. It is a standard deployment that we are used to. During our initial validations and evaluations of
total), the
cluster, we are reaching write speeds between 250-300MB/s, which would be the ballpark for this kind of setup for HDDs with the NVME as Rocks.db cache (in our experience). However, the issue is the reading process. While reading, we barely hit the mark of 100MB/s; we would expect at least something similar to the write speed. These tests are being performed in a pool with a replication factor of 3.
We have already checked the disks, and they all seem to be reading just fine. The network does not seem to be the bottleneck either (checked with atop while reading/writing to the cluster).
Have you guys ever encountered similar situations? Do you have any tips for us to proceed with the troubleshooting?
We suspect that we are missing some small tuning detail, which is affecting the read performance only, but so far we could not pinpoint it. Any help would be much appreciated :) _______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io
-- +4916093821054 _______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io
participants (4)
-
Arvid Picciani
-
Janne Johansson
-
Marc
-
Rafael Weingartner