Steve Modica
Forum Replies Created
-
[John Davidson] “Does the manual show how to configure correctly with OS X Server / Mountain Lion?”
You just have to load the Gigabit driver from our download page (best to do that before installing the card), then power down and put the card in. The ports will show up just like mac ports in the System Preferenes->Network pane. We have link agg instructions in our knowledge base with screen shots.
Steve
Steve Modica
CTO, Small Tree Communications -
Small Tree cards all work on Mountain Lion and with hotplug/unplug over thunderbolt
Steve Modica
CTO, Small Tree Communications -
Depending on your timeframe you might want to wait another quarter. It looks to me like FCoE is starting to heat up. It doesn’t really have a performance advantage over iSCSI right now, but the offloads are there and the CPU utilization is lower. We got some amazing IOP counts with FCoE and our linux target code.
As I understand it, FCoE targets are supposed to be arriving in Q3/4. You can also get an FCOE switch and plug in existing FC targets and it’ll convert them to FCoE.
Steve
Steve Modica
CTO, Small Tree Communications -
You might try the Small Tree Initiator. We just got MPIO working and I’d like to see if it works with your target anyhow. (Not sure if we posted the driver but you can email me and I’ll send it).
I don’t think MPIO is going to help you. It will increase the bandwidth to the thing, but you could do that with 10Gb too (assuming it supported 10Gb). The issue will be the spindles. For $499 you could get maybe 1 spindle 🙂 That’s not enough for video latencies.
People go around thinking that the network is the problem or that ethernet is the problem. Ethernet ROCKS. Ethernet is superfast. Ethernet has flow control and the gigabit port on your Mac Book Pro can *kick your internal spinning drives paltry little 40MB/sec ASS*, yet, people think Ethernet is slow.
So, the issue is that disk drives are slow. Disk drives need to catch up to 10Gb now. (1GB per sec!) Bring it on.
Steve
Steve Modica
CTO, Small Tree Communications -
Hi Chris
Sorry to be slow in responding. We do a lot of military work and I was working on an RFI response for a govt agency these last few days.
The size and performance of a server is of course determined by the load you anticipate and the speed, performance and number of clients you plan to attach. I’ve been asked about a zillion times for specifics, but there are no hard fast rules.
Here are some of mine:
1. The server should be as fast as the fastest client or *at least* of the same CPU generation as the fastest client
We are dealing with 3 broad generations right now, soon to be 4. Old Mac Pros, Nehelam, Westmere and the new guys (Sandy bridge/Ivy Bridge).
You can’t expect a first gen Intel Mac Pro to handle the data coming from a Westmere client.Does this mean it won’t work? No, but it does mean that if you decide to do a very taxing render to the server, you might cause other clients to slow down and drop frames. So if you’re asking me to spec things, I have to consider worst case. Whereas a lot of people looking to save money would rather deal with the occasional slow down.
2. More memory is better
It used to be 1 GB of memory per client was about right. I now tell people 2 GB of memory. I would prefer even more.
The more memory you have on the server, the more the server can cache and the more efficient the IO to the RAID can be. If you are low on memory, tiny IOs like directory queries will hit the raid leading to poor performance.
3. 10Gb needs the fastest macs.
As things are today, Only the westmere systems can really drive a significant number of 10Gb ports. So if you want to do 10Gb to clients, make sure the machines with 10Gb cards have fast cpus. The Westmere system has the memory controller inside the cpu which makes a huge difference. (Sandy bridge moves the Northbridge inside the CPU which will be *another* huge increase in throughput.)
4. Get the slots right
Big IO cards need big slots. The server has 2 16X slots and one needs to be for the RAID and one needs to be for the network card. The graphics card will have to move.
5. Tune the system correctly.
You are going to be using a deskside workstation as a high throughput device. It’s not tuned for that. Apple doesn’t consider that you’ll have 200,000 open files or that you’ll have 1GB/sec moving from disk to the network all day. So you need lots of nbufs, clusters, vnodes, open file descriptors etc. (We tune all systems that have our storage attached). One simple way to get some of this tuning is to load OS X server. They do some of it.
Steve
Steve Modica
CTO, Small Tree Communications -
10Gb uses NAS protocols (like AFP and Samba) and files are accessed as files off of a server. It’s easier to setup, simpler to understand and has only one server as a point of failure (which you can make redundant via LACP, dual power etc).
Fibre channel and iSCSI are not shared. The idea is that clients access the storage directly with a metadata server playing traffic cop. Everyone is diddling with the inode table via the metadata server. We’ve set these up with our iSCSI to test performance and compatibility (iSCSI 10Gb is faster than FC BTW). It’s complicated. You have to have lots of LUNS to stripe together plus a metadata LUN that’s mirrored.
My decision point would be dependent on the number of clients and bandwidth required. If you can get away with NAS, do NAS. It’s much easier to deal with (and cheaper). If you need more than you can get from a single server, consider 2 or 3 servers.
Steve
Steve Modica
CTO, Small Tree Communications -
I can comment on our testing:
Since Thunderbolt is 4X and our 10Gb cards are 8X, we have noticed a slight reduction in performance. We don’t get Line rate on tests, we get about 80% of that.
Hotplug and unplug are pretty amazing. Once we get it right, you can unplug the thing when stuff is going on. The activity will hang (and eventually timeout), but if you reconnect it recovers. The machine doesn’t panic.
Steve Modica
CTO, Small Tree Communications -
[Cesar Cabrera] “Thanks a lot! But we want to put the clients on 10gige and keep the the fiber protocol only between the xserve and galaxy. The clients we would like to have them on 10gige network.”
I think you’d have to ask Rorke in that case. I’m not sure what they allow (in terms of protocols and performance) from their Galaxy box.
Steve Modica
CTO, Small Tree Communications -
[Cesar Cabrera] “It is possible to install a small tree or atto 10gig E to already running Xsan with a Rorke Galaxy 3g and change the network from fiber to ethernet 10gig and leave the Galaxy and the server in fiber?”
Not sure of the layout here. If the clients currently have FC cards, you’d have to switch them over to FCoE and purchase an FCoE switch to handle the conversion between the MDCs and the storage’s FC ports. Cisco Nexus switches do this (that’s what we use internally for development).
Steve
Steve Modica
CTO, Small Tree Communications -
You didn’t say anything about your storage. CPU cycles and memory accesses are measured in microseconds, storage accesses are usually milliseconds. So if things seem sluggish, start there.
Steve
Steve Modica
CTO, Small Tree Communications