Steve Modica
Forum Replies Created
-
[Bernard Newnham] “I must admit I didn’t know the name Westmere, but Wikipedia tells me it’s just another Intel processor architecture – a new one every week folks. This one is used for making Core i7 processors and successors, since that’s what they’re making currently. I’ve made a hackintosh on both P35 and P55 chipsets now – the operating system doesn’t appear to be that picky, given a little help from it’s Linux/Unix friends, so it will probably work on others. Macs are much less special than Apple would like you to think….”
Westmere is the server atchitecture. i3, i5 and i7 are the desktop architectures. Both offer the integrated memory controller, but the westmere is 3 channel. I think the i3-i7 are two channel (it makes a difference for networking).
Steve Modica
CTO, Small Tree Communications -
[Bernard Newnham] “Not sure what you mean by that. All that matters is that FCP – or whatever application – works on the machine. It’s worked twice on two different motherboard/CPU combinations for me, so it probably works on lots more. In fact tonymac https://tonymacx86.blogspot.com/ thinks his method of hackintosh build will work on lots of systems. You do have to remember that OSX is secretly – under the pretty interface – just another version of Linux/Unix, which is why people can make the thing work.”
FCP and Aja and a lot of things things FCP will interact with use drivers. Those drivers are built and tested on Macs using the specific Nehelam or Westmere chipsets Apple is using. Using some other system with some other PCIE, USB, graphics or other chipset may lead to unexpected issues.
Even if you match all the hardware, you can still run into firmware/initialization issues because something is brought up slightly differently with firmware XYZ.
As a hobby, I think it’s a great idea. I’d love to have OS X running on lots of old hardware we have kicking around.
Steve Modica
CTO, Small Tree Communications -
I think it’s crazy depending on what you’re trying to do.
I can’t use them because if I got a driver built and working on one, I couldn’t be sure it worked on a mac, so I’d have to get macs anyhow.Steve Modica
CTO, Small Tree Communications -
We would do this:
Open a terminal and run top. While that’s going, hit your edit to tape thing and see if something comes up that’s using a lot of cpu (or at least actively running. Take a note of it’s PID (Process ID) on the left.Once you know what that is, kill top and run this:
sudo dtruss -p PID
(you will have to enter and administrator password)If the app is reading or writing or doing anything with the kernel (going out on the net etc), it will make tons of system calls and you will see them scrolling by. I would expect lots of reads and writes of one form or another.
You can stop it by hitting “control C” in the terminal.
I imagine it’s creating some gigantic index file or image file for tape and it takes a long time.
Once you’ve seen what happens on start up, try killing the “edit to tape” thing and see what happens. I expect it’s deleting a bunch of stuff.
Steve Modica
CTO, Small Tree Communications -
[Eric Jurgenson] ”
So you are saying that 4-1G connections between the server and the switch (link aggrigation mode – NAS configuration) is likely to outperform a 10G connection? Is this just for certain switch models?I was thinking of going with an SMC-8926EM managed switch with a 10G module for my server connection. Any comments on this particular switch regarding 10G to 1G flow control and other issues?”
When a 10Gb port is properly configured (on both sides) and flow control is working, it should be able to satisfy 5-6 Gigabit ports doing substantial work (think 50-60MB/sec. Beyond that and you’ll be getting memory allocation glitches and things)
If you are only planning on deploying a single 10Gb port, you may as well just do Gigabit with a 6 port card.
I don’t know anything about that specific switch. SMC uses broadcom and our previous experience with them is that flow control was disabled due to a head of line blocking bug. Whether they’ve fixed that I don’t know.
All of these vendors are really evasive about Flow control. They call it out in their specs, but then bury something in the manual explaining that it doesn’t work right 🙁
Steve
Steve Modica
CTO, Small Tree Communications -
CPU is usually not the problem. In fact, when we’re spec’ing out servers for people, we usually have them buy the slowest clock rate. Memory and disk performance are the big bottlenecks for most people.
Unless you are running lots of things on the system at once, I’d consider disk first.
The eSATA raids I’ve tested (and it wasn’t many) were not very good.Steve Modica
CTO, Small Tree Communications -
When you are doing things locally, you get the benefit of the systems local cache, so a lot of latency issues with a single disk are hidden. That being said, a single disk, or even two striped disks aren’t going to be fast enough for much sustained streaming. 8 is the fewest spindles we’ll go with.
This isn’t about the sustained IO rate of the devices. It’s about the ability of the device to keep a low latency response time when you’ve got three file pointers going at once (audio left and right and the video itself). Obviously more things on the time line leads to more file pointers.
The fastest I’ve seen FW800 go was 64MB/sec. I haven’t tried it with an SSD.Steve Modica
CTO, Small Tree Communications -
[Eric Cox] “Latency is going to suffer with larger stripe sizes. The is simply because the head has to read the full stripe before moving onto the next batch of sectors. Large stripes are typically used where you need increased throughput on large linear reads.”
I’m going to respectfully disagree in our context of video editing. The applications all issue 4MB reads. The days of reading a frame aligned IO are gone. The App is just sucking in chunks as fast as it can. Further, the OS/Filesystem code is aggregating those IOs. For example, if you issue a number of 1MB linear IOs, the filesystem code will aggregate those up to 32MB (in 10.6).
What this means in practice is you want a full stripe size that matches that. Otherwise, you increase the number of stripes you need to read to get the data off (latency ends up going up with overhead, not down).
The notion of a stripe at 128k or 512k is just stupid at these IO sizes, which is why striping stinks so bad.I believe most of the modern raid controllers will also service small IOs with a partial stripe read. (This is a guess, but I know they don’t all check parity on every IO either).
To summarize, with the large, aligned IOs the apps are issuing, bigger is usually better.
Steve Modica
CTO, Small Tree Communications -
[Bruce Little] “How about RAID 50?”
Any of the 0 stripe setups are problematic.
With two RAID5s, if either has a problem, you are performance limited. So you double your chances for problems. (or at least performance limited periods)Secondly, the striping drivers that are out there all suck. They give each stripe 256k or 512k and that’s very little relative to what the raids really want. They do best with larger IOs. In Small Tree’s testing, any striping makes lantency much worse.
Steve
Steve Modica
CTO, Small Tree Communications -
[Eric Cox] “Don’t do RAID 5 on multi-terabyte volumes. Even if the array has hot spares available this is asking for issues. The problem being that the potential for a misread while rebuilding from parity is now so great that rebuilds can fail. The read error rate has remained constant, but we have scaled disk sizes up dramatically from the days of raid 5’s inception.”
The choice depends on the drives you use. Typical SATA drives (even ES drives) have an error rate of 1 in 10^14. SAS drives are lower density and offer 10^16. The SATA drives we use now are 10^15 blocks. That works out to about 1 read error in 120TB. I draw the line at around 10 drives to go to RAID6.
Steve Modica
CTO, Small Tree Communications