Steve Modica
Forum Replies Created
-
Hey Bob.. How did you ever get that 10Gb stuff working so fast? I forget 🙂
Steve Modica
CTO, Small Tree Communications -
[Morten Ranmar] “Could I not install an extra ethernet card in the clients, and connect each via 2 ethernet cables to the 6-port Small Tree card, and get higher throughput?”
This won’t work reliably due to the random nature of LACP. It does socket balancing. So the sockets get randomly assigned between the ports. The algorithm on a mac uses a hash of source and destination mac addresses plus outbound port number (which changes).
In a best case scenario, the inbound file (your source) will come in on one port and the outbound (your rendered file) will go out on the other port. However they could randomly end up on the same port. There’s no way to control it. (I’ve never tried mounting on one port with AFP and one with SAMBA and then treating them as separate servers. That would be interesting, but ugly.)
A better answer would be to put a 10Gb 6 port card in the server and run 10Gb to the clients.
Steve Modica
CTO, Small Tree Communications -
[Bob Zelin] “As far as creating the “bottleneck”, I totally disagree with the super smart guys that have responded to you. The typical user is not using the Mac server as a high speed server. You are doing normal file serving (possibly with regular ethernet), and things like FCP can’t take advantage of the faster machines. If you have a new 12 core (for example), you can make this a workstation, where you do DEDICATED LOCAL WORK with Adobe CS5 After effects (for example) and get a hi performance local machine, that you will “occationally” send files to the “old dog” file server from 2008.”
If what Bob says is true (that this is not going to be used for editing in place) then I agree. You can use the slower machine.
However, people drag and drop all the time and if they do it in such a way that the data flows through the faster client, it will botteneck the server in a bad way. (It’ll hit other people on the network too)Steve Modica
CTO, Small Tree Communications -
[Morten Ranmar] “I bought the new (xServe replacement) MacPro Server to build a SAN setup (3 clients) – but am considering swapping it for a 2008 MacPro, in order to use the new faster machine for a workstation, and the older machine as a Server. Does it sound reasonable?”
I wouldn’t do this. The Mac Pro’s today have an integrated memory controller. It’s something like 4X faster than the pre-nehelam systems.
In the networking world, memory bandwidth is everything. All packets get memcopied.
The upshot is an “old” mac pro might sustain 30MB/sec bidirectional on a render, whereas the “new” Nehelams will do about 60MB/sec bi directional.
The result is you beat the heck out of the gigabit port and it has to flowcontrol a lot.
If you are going to create a congestion point, create it on the clients.Steve Modica
CTO, Small Tree Communications -
The Aja test is not a good representation of FCP traffic. It uses a smaller IO size and it’s not Async IO either. FCP does larger IOs in parallel.
Steve Modica
CTO, Small Tree Communications -
Mark my words… EtherNOT never wins.
SCSI, FC, Token Ring, Fddi, ATM, GSN, HIPPi-800, Infiniband, x.25 and whatever else you want to mention all fall away as ethernet comes along and whomps them. There are too many ethernet ASICs out there in the world doing real work. What’s the one chip architecture that’s on *all* motherboards? The one that has a higher capture rate than even Intel’s CPU? Ethernet…To quote another guy who’s name I can’t remember “I don’t know what the future of networking will look like, but I know it will be called Ethernet” (OpenFabric and iWarp are basically an attempt to get Infiniband functionality and virtualization on to Ethernet)
Steve Modica
CTO, Small Tree Communications -
Before Small Tree was doing shared storage and ethernet, we were doing infiniband for the PowerPC Xserve. We had released the product and were selling it for about 3 months when Apple announced they were ditching PPC. There was no Xserve replacement at the time and would not be for about 1.5 years. Since PPC was dead, people stopped buying. We had to scramble into another business or go under. I learned my lesson!
Of course, doing Inifiniband drivers for Mac helped us learn a lot of about how it all worked.
Steve Modica
CTO, Small Tree Communications -
[Nate Cooper] “Same with small tree, Small Tree is awesome hardware, that I personally incorporate with Promax ShareMAX just as you incorpoorate it with Maxx Digital Final Share. Traditionally ST components are implemented through VARs to be part of a complete solution.”
Small Tree sells ST RAID and ST Mobile (both part of Granitestor). We have complete solutions.
What’s nice about ours is that you get our network cards and our storage. We have it covered end to end and we understand how the kernel works on all of it since we either write or heavily tweak all the drivers.
Steve Modica
CTO, Small Tree Communications -
Keep in mind that Gigabit and 10Gb are like plumbing. You need bigger pipes to feed the smaller ones. Up til now, everyone has been 1Gb and 10Gb hasn’t been very widely deployed as “the bigger pipe”. It’s been lots of link aggregation.
As 10Gb starts to hit the desktop, the first “big pipes” will be cards like our 6 port 10Gb card. I would love to have 100Gb to follow up with.
Lightpeak is going to be a firewire/USB kind of thing. It’s not going to supplant Ethernet. There’s Ethernet and EtherNOT and EtherNOT never wins.
Steve Modica
CTO, Small Tree Communications -
[Wayne Carey] “Exactly, Chris. This my point. Not all businesses live in fantasy world of all up to date software. Some of us live in the real world where our desktop systems are still running Windows XP and Server 2003. Alot of the corporate world is afraid of Windows 7 and Server 2008. Apple should have some understanding of this.”
I was a sysadmin at Ford for a number of years. I was also at SGI working in engineering and acting as part of the release team (approving changes etc)
When you’ve got hundreds of people working on something, you can only release in controlled periods. Otherwise, it gets crazy as different changes step on each other. Everyone has a schedule for when they’ll have their new feature or bug fix finished and they have to have some sort of framework to schedule against.
A lot of sysadmins see danger in change so they stay with what they have as long as they can. I think that makes sense for one release iteration, but more than one is asking for trouble.
For example, I can see people sitting at 10.5.8. If they are production and in the middle of something, that’s OK. However if they are at 10.4.11, that’s asking for trouble. Any systems running an OS that old should be used as appliances only, in situations that are known to be safe and working. They should not be redeployed because they will likely be broken.
Steve Modica
CTO, Small Tree Communications