XenServer hyperconverged
-
@olivier Here are my results (Windows Server 2008, LVM, Raid 10, 8x 15K spinning rust) --
Sequential Read (Q= 32,T= 1) : 416.270 MB/s
Sequential Write (Q= 32,T= 1) : 412.617 MB/s
Random Read 4KiB (Q= 32,T= 1) : 14.298 MB/s [ 3490.7 IOPS]
Random Write 4KiB (Q= 32,T= 1) : 17.564 MB/s [ 4288.1 IOPS]
Sequential Read (T= 1) : 321.305 MB/s
Sequential Write (T= 1) : 273.068 MB/s
Random Read 4KiB (Q= 1,T= 1) : 1.218 MB/s [ 297.4 IOPS]
Random Write 4KiB (Q= 1,T= 1) : 12.264 MB/s [ 2994.1 IOPS]Test : 1024 MiB [C: 75.9% (56.9/75.0 GiB)] (x5) [Interval=5 sec]
-
@Danp Your storage is on your host right?
-
@olivier Yes, local storage.
-
@Danp Thanks!
-
Okay, so after just few days of technical experiments, here is the deal.
Context
- 2x XS7 hosts, installed directly on 1x Samsung EVO 750 (128 GiB) each
- dedicated 1Gb link between those 2 machines (one Intel card, the other is Realtek garbage)
Usually, in a 2 hosts configuration, it's not trivial to avoid split-brain scenarios.
In a very small setup like this (2 hosts only with few disk space), you'll expect the overhead to be the worst possible regarding the proportion of resources. But will see it's still reasonable.
Current working solution
A shared file storage (thin provisioned):
What's working
- data replicated on both nodes
- fast live migrate VMs (just the RAM) between hosts without a NAS/SAN
- very decent perfs
- "reasonable" overhead (~2GiB RAM on each Node + 10GiB of storage lost)
- scalable up to the max pool size (16 hosts)
- killing one node and other VMs on the other host will still work
- using XenServer HA on this "shared" storage to automatically bring back to life VMs that were on the killed node
- no split brain scenario (at least during my tests)
- no over complicated configuration on hosts
Overhead
- RAM overhead: <5GiB RAM on 32GiB installed
- Storage overhead: lost around 9GB of disk space per host
Obviously, in case of using large local HDDs, storage overhead will become negligible.
Scalability
In theory, going for more than 3 nodes will open interesting perfs scalability. So far, it's just replicating data, but you can also spread them when you have 3+ nodes.
Perfs
I'm comparing to a dedicated NAS with ZFS RAID10 (6x500GiB HDDs) with 16GiB of RAM (very efficient cache for random read/write) with semi-decent hardware (dedicated IBM controller card), on a NFS share.
ZFS NAS XOSAN diff Sequential reads 120 MB/s 170 MB/s +40% 4K reads 9.5 MB/s 9.4 MB/s draw Sequential writes 115 MB/s 110 MB/s -5% 4k writes 8.4 MB/s 17 MB/s +200% As you can see, that's not bad.
Drawbacks
- right now, it's a fully manual solution to install and deploy, but it could be (partly) automated
- it's a kind of "cheating" with XAPI to create a "shared" local file SR (but it works ^^)
- XS host can't mount the share automatically on boot for some reasons. So I'm currently finding a way to do that correctly (maybe creating a XAPI plugin?)
- you'll have to deploy 2 or 3 rpm's on Dom0, but the footprint is pretty light
- it will probably (very likely in fact) work only on XS7 and not before
- the only clean way to achieve this is to have SMAPIv3 finished. Until then, we'll have (at XO) to glue stuff in the best way we could to provide a correct user experience.
Conclusion
It's technically doable. But there is a mountain of work to have this in a "one click" deploy. I'll probably make a closed beta for some XOA users, and deploy things semi-manually to validate a bit the concept before spending to much time scaling something that nobody will use in production for some reasons (interest, complexity, etc.).
-
@olivier
Very nice work.
I'm claiming my place on the beta line. -
@olivier
Any drawback while using the default storage type LVM? -
@black3dynamite I'm not sure to understand the question.
So far, the "stack" is:
- Local Storage in LVM (created during XS install)
- on top of that, filling it by a big data disk used by a VM
- the VM will expose this data disk
- XenServer will mount this data disk and create a file level SR on it
- VMs will use this SR
It sounds like a tons of extra layers, but that's the easiest one I found after a lot of tests (you can see it as a compromise between modifying the host too deeply to reduce the layers VS not modifying anything into the host but have more complexity to handle on VM level). You can consider it as an "hybrid" approach.
Ideally, XenServer could be modified directly to allow this (like VMWare do with VSAN), and expose the configuration via XAPI.
I think if we (XO project) show the way, it could (maybe) trigger some interest on Citrix side (which is only into XenDesktop/XenApp, but hyperconvergence even make sense here)
-
@olivier
The question is based on you using EXT (thin) shared storage instead of LVM (thick) for XenServer. -
@black3dynamite I can't use LVM because it's block-based. I can only work with file level backend.
I did try to play with blocks, performance was also correct (a new layer so a small extra overhead). But I got a big issue in certain cases. Also, it was less scalable in "more than 3" hosts scenario.
-
@olivier Sounds promising. Can you elaborate on how adding additional overhead of XOSAN would yield an increase in performance of 40% / 200% ?
-
@Danp That's because despite the added layers, we are making some operations locally and also on a SSD (vs the existing NAS where everything is done remotely)
It's more indicative than a real apple to apple benchmark. The idea here, is to show the system would be usable, performances are not the main objective here.
-
Blog post available: https://xen-orchestra.com/blog/xenserver-hyperconverged/
-
I also updated the benchmarks with a Crystal Diskmark working on a 4GiB file (avoiding ZFS cache). Difference of performance is now huge, so the impact of the replication is not that bad at the end.
-
@olivier said in XenServer hyperconverged:
I also updated the benchmarks with a Crystal Diskmark working on a 4GiB file (avoiding ZFS cache). Difference of performance is now huge, so the impact of the replication is not that bad at the end.
Awesome! This is all very exciting.
-
Really looking forward to playing with this
-
Just bumping up. Hopefully, Olivier will have some favorable update.
EDIT: corrected olivier...my apologies
-
Well, sort of
This command is now working:xe sr-create name-label=XOSAN shared=true content-type=user type=xosanStill a lot of stuff between this and a viable product, I'm in the middle of testing the solution in terms of resilience and overhead. I need also a layer of glue to at least semi-automate the deployment of XOSAN on a pool, otherwise I'll spent to much time doing it manually for each beta tester ^^
Anyway, developing a storage on current storage architecture of XenServer is really a pain. Eager to see SMAPIv3 in action
-
Thanks @olivier
Just finished reading your blog with response to comment referring to January for beta...I think it's too early in January. Getting excited here just by reading. Nice idea/product, btw.
More power!
-
I still have 27 days to stay in track
