Hi Angelo,
Just some thoughts to consider from our experience with similar setups:
1. Use Proxmox instead of VMWare, or anything KVM based. These VMs can
consume Ceph directly, and provide the same level of service (some may
say better) for live ,migration, hyperconvergence etc. Then you run
Windows VMs in KVM, bring RBD storage to them as virtual disks and share
out as needed.
2. Use NFS - all modern Windows OSs support it. You can use any NFS
gateway you like, or set up your own machine or cluster (which is what
we did with Storcium) and export your storage as needed.
3. If you must use VMWare, you can present datastores via NFS as well,
this has a lot of indirection but is easier to manage.
--
Alex Gorbachev
ISS Storcium
https://www.iss-integration.com <https://www.iss-integration.com>
On Thu, Apr 27, 2023 at 5:06 PM Angelo Höngens <angelo@xxxxxxxxxx
<mailto:angelo@xxxxxxxxxx>> wrote:
Hey guys and girls,
I'm working on a project to build storage for one of our departments,
and I want to ask you guys and girls for input on the high-level
overview part. It's a long one, I hope you read along and comment.
SUMMARY
I made a plan last year to build a 'storage solution' including ceph
and some windows VM's to expose the data over SMB to clients. A year
later I finally have the hardware, built a ceph cluster, and I'm doing
tests. Ceph itself runs great, but when I wanted to start exposing the
data using iscsi to our VMware farm, I ran into some issues. I know
the iscsi gateways will introduce some new performance bottlenecks,
but I'm seeing really slow performance, still working on that.
But then I ran into the warning on the iscsi gateway page: "The iSCSI
gateway is in maintenance as of November 2022. This means that it is
no longer in active development and will not be updated to add new
features.". Wait, what? Why!? What does this mean? Does this mean that
iSCSI is now 'feature complete' and will still be supported the next 5
years, or will it be deprecated in the future? I tried searching, but
couldn't find any info on the decision and the roadmap.
My goal is to build a future-proof setup, and using deprecated
components should not be part of that of course.
If the iscsi gateway will still be supported the next few years and I
can iron out the performance issues, I can still go on with my
original plan. If not, I have to go back to the drawing board. And
maybe you guys would advise me to take another route anyway.
GOALS
My goals/considerations are:
- we want >1PB of storage capacity for cheap (on a tight budget) for
research data. Most of it is 'store once, read sometimes'. <1% of the
data is 'hot'.
- focus is on capacity, but it would be nice to have > 200MB/s of
sequential write/read performance and not 'totally suck' on random
i/o. Yes, not very well quantified, but ah. Sequential writes are most
important.
- end users all run Windows computers (mostly VDI's) and a lot of
applications require SMB shares.
- security is a big thing, we want really tight ACL's, specific
monitoring agents, etc.
- our data is incredibly important to us, we still want the 3-2-1
backup rule. Primary storage solution, a second storage solution in a
different place, and some of the data that is not reproducible is also
written to tape. We also want to be protected from ransomware or user
errors (so no direct replication to the second storage).
- I like open source, reliability, no fork-lift upgrades, no vendor
lock-in, blah, well, I'm on the ceph list here, no need to convince
you guys ;)
- We're hiring a commercial company to do ceph maintenance and support
for when I'm on leave or leaving the company, but they won't support
clients, backup software, etc, so I want something as simple as
possible. We do have multiple Windows/VMware admins, but no other real
linux guru's.
THE INITIAL PLAN
Given these considerations, I ordered two identical clusters, each
consisting of 3 monitor nodes and 8 osd nodes, Each osd node has 2
ssd's and 10 capacity disks (EC 4:2 for the data), and each node is
connected using a 2x25Gbps bond. Ceph is running like a charm. Now I
just have to think about exposing the data to end users, and I've been
testing different setups.
My original plan was to expose for example 10x100TB rbd images using
iSCSI to our VMware farm, formatting the luns with VMFS6, and run for
example 2 Windows file servers per datastore on that with a single DFS
namespace to end users. Then backup the file servers using our
existing Veeam infrastructure to RGW running on the second cluster
with an immutable bucket. This way we would have easily defined
security boundaries: the clients can only reach the file servers, the
file servers only see their local VMDK's, ESX only sees the luns on
the iSCSI target, etc. When a file server would be compromised, it
would have no access to ceph. We have easy incremental backups,
immutability for ransomware protection, etc. And the best part is that
the ceph admin can worry about ceph, the vmware admin can focus on
ESX, VMFS and all the vmware stuff, and the Windows admins can focus
on the Windows boxes, Windows-specific ACLS and tools and Veeam
backups and stuff.
CURRENT SITUATION
I'm building out this plan now, but I'm running into issues with
iSCSI. Are any of you doing something similar? What is your iscsi
performance compared to direct rbd?
In regard to performance: If I take 2 test windows VM's, I put one on
an iSCSI datastore and another with direct rbd access using the
windows rbd driver, I create a share on those boxes and push data to
it, I see different results (of course). Copying some iso images over
SMB to the 'windows vm running direct rbd' I see around 800MB/s write,
and 200MB/s read, which is pretty okay. When I send data to the
'windows vm running on top of iscsi' it starts writing at around
350MB/s, but after like 10-20 seconds drops to 100MB/s and won't go
faster. Reads are anywhere from 40MB/s to 80MB/s, which is not really
acceptable.
Another really viable and performant scenario would be to have the
Windows file servers connect to rbd directly with the windows rbd
driver. It seems to work well, it's fast, and you don't have the
bottleneck that the iscsi gateway creates. But I see this driver is
still in beta. Is anyone using this in production? What are your
experiences? We would miss out on the separation of layers and thus
have less security, but at the same time, it really increases
efficiency and performance.
And if I use rbd, then vmware won't see the storage, and I cannot do
an image backup using veeam. I could of course do backups of the rbd
images, using tools like restic or backy to rgw running on the second
cluster with immutable buckets. What are your experiences? Is it easy
to do differential backups of lots of 50TB rbd images? Change rate is
usually like 0.005% per day or something ;)
By the way, we also thought about CephFS, but we have some complex
stuff going on with extended ACL's that I don't think will play nice
with CephFS, and I think it's a lot more complex to backup CephFS than
block images.
If you made it here, thank you for your time! I hope you can share
thoughts on my questions!
Angelo.
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
<mailto:ceph-users@xxxxxxx>
To unsubscribe send an email to ceph-users-leave@xxxxxxx
<mailto:ceph-users-leave@xxxxxxx>
