[Steve Modica]

In the olden days, some apps and OS’s used signed pointers and some didn’t. If you have a signed pointer, you only get 2^31 bits (2GB). I thought FCP 7 was using signed pointers (like SGI used to). Clearly they aren’t (like Solaris).

I think that’s what OP is hitting tho. The 4GB limit.

Steve Modica
CTO, Small Tree Communications

[Steve Modica]

FCP 7 is a 32 bit app. It can only address 2GB of memory (using signed pointers). So once it has filled up 2GB, it’s out of memory.

Steve Modica
CTO, Small Tree Communications

[Steve Modica]

I was playing with this not long ago. I think the issue is configuration on the target side. (this is freeBSD I was playing with). I saw similar problems.
FreeBSD 10 builds the iSCSI target code into the kernel, so everything changes anyhow.

Steve Modica
CTO, Small Tree Communications

[Steve Modica]

[Eric Hansen] “- Barefeats got the Areca 8050 up to 900MB/s. it stayed there in a few different configs (RAID0, RAID5, etc), so that looks to more or less be the limit for TB1. Now, I may be totally wrong here, but with TB1, doesn’t that mean the storage is using a single 10Gb lane of the 2 available on the port? would that mean that a 10GbE card on the same TB1 port (daisy chained) would use the second 10Gb channel? thus the drive would still keep hitting 900MB/s while the 10GbE card will go as fast as the 10Gb channel allows?

– in your Adobe demo, you say “Thunderbolt 1 forced me to use 2 10Gb ports to go that fast”. does that mean you used two Thundernet boxes and bonded them in System Prefs to hit 1GBs? if so, does that mean 10GbE bonding works differently than 1GbE bonding, where you can’t actually increase speed with bonding unless there’s something “smart” controlling it, like Avid ISIS?”

On TB1, the graphics channel is not available for IO. (It is on TB2).
I think the theoretical limit for TB1 is 1GB/sec (which a storage controller can approach), but for 10Gb and non-optimal IO (like many users reading from the storage) you’ll see much lower numbers. Areca loves to post those high BW numbers 🙂

As for the 2 TB ports at NAB, I didn’t use bonding. I used iSCSI and MPIO. That allowed me to ping pong scsi commands between 10Gb cards (each on its own thunderbolt port). LACP would have stuck me with a single socket using only one of the ports.

Steve Modica
CTO, Small Tree Communications

[Steve Modica]

It’s been my experience that iSCSI is much faster than AFP/Samba. AFP and Samba are shared protocols and have to do a lot of stat’ing of directories before they do anything (in fact, as you copy files, they are constantly double checking that you are still allowed to copy the file).

iSCSI is not a shared protocol. So it doesn’t do any of this. So the access is either volume shared or SAN/metadata controller based. iSCSI can generally hit line rate on 1Gb and 10Gb. (This is exactly why I was using it at NAB to do my 1GByte/sec Adobe demo, although as you might recall, Thunderbolt 1 forced me to use 2 10Gb ports to go that fast)

I think the problem in this scenario is going to be the Thunderbolt port. I don’t think it’s capable of more than 600MB/sec. Divide that between the two devices (RAID and network) and your at 300MB/sec. Divide that amongst 6 Gigabit ports and you’re at 50MB/sec.

Steve

Steve Modica
CTO, Small Tree Communications

[Steve Modica]

My experience with AFP is that it was easier to get larger IO sizes out of it to make the server storage happier. Samba always generated smaller IOs.

That being said, Samba did a much better job scaling. One users activity impacted other users a lot less. In fact, I have a bug going on right now where Premiere is stat’ing a ton of files and causing AFP to hit very high CPU numbers. This takes out other users. Samba sails right through it. The individual user sees the same issue, but the other users don’t get impacted. Samba scales better and we can get much larger aggregate stream counts with Samba.

One other advantage of Samba is it’s VFS layer. Samba gives vendors that ability to insert modules within the data path. If you look at a unix system with samba installed and dig around, you’ll find a vfs directory (that comes with samba). In there, you’ll see all these vfs modules:

titaniumz:/usr/local/lib/samba/vfs# ls
acl_tdb.so commit.so full_audit.so scannedonly.so time_audit.so
acl_xattr.so crossrename.so linux_xfs_sgid.so shadow_copy.so weird.so
aio_fork.so default_quota.so netatalk.so shadow_copy2.so xattr_tdb.so
audit.so dirsort.so preopen.so smb_traffic_analyzer.so zfsacl.so
cacheprime.so expand_msdfs.so readahead.so streams_depot.so
cap.so extd_audit.so readonly.so streams_xattr.so
catia.so fake_perms.so recycle.so syncops.so

These all do different things. They are all open source. Many of them are written by specific vendors to highlight their platform (IBM, SGI, Isilon).

Some of them deal with mundane permissions or file naming issues. Some of them attempt to deal with big IO performance issues (cacheprime and readahead for example).

What’s wonderful about this is that vendors can insert code here to improve their own performance.

The one thing we’ll lose that a lot of apple customers liked was “ignore ownership on this volume”. I don’t think Samba has anything like that other than user squashing. Perhaps we’ll have to do that.

Steve

Steve Modica
CTO, Small Tree Communications

[Steve Modica]

Hi John
For print, usually the tuning requirement is just to go as fast as possible. The apps behave differently, but there’s usually no realtime component (IE you won’t drop frames. You just want your file copies to be fast).

The Small Tree team was doing this back with SGI in the late 90s and early 00s before becoming Small Tree, so we have a lot of experience in print. We also are using ZFS now so we have huge expansion potential without doing raid rebuilds. Check out the TitaniumZ line. We showed it off at NAB. We have the portable Z5 all the way up to the Z16.

Steve

Steve Modica
CTO, Small Tree Communications

[Steve Modica]

Wifi isn’t inherently bad. It’s just that anything you can do via wireless you can probably do better over wired. So when they finally get around to full duplex, 10Gb wireless, there will be 100Gb Ethernet. Maybe all the codecs will be tiny so it won’t matter, but more likely, someone will be shooting 50k video and they’ll need 20GB/sec to play it out.

Wired has the advantage of being shielded and being a better conductor of high frequency signals.

Steve Modica
CTO, Small Tree Communications

[Steve Modica]

Hi John

You should take a look at Small Tree TitaniumZ-8. You can start off with a relatively small amount of ZFS storage and we can grow that on the fly if you need to without rebuilding the array. We’re kernel guys so when it comes to figuring out how an app works, we’re good at it. TitaniumZ-8 is running FreeBSD (which is like OS X’s big brother) so we can use a lot of the same tools and performance enhancements to get things working very smoothly.

Steve Modica

Steve Modica
CTO, Small Tree Communications

[Steve Modica]

I have to say, Adobe has been the most amazing partner since “the schism” that was FCP X.
That particular quicktime bug has been around for a long time and we’ve worked very hard to fix it (I sent lots of traces to apple etc). Adobe is the only vendor that actually went out and fixed it.

So to summarize (and I’ll be blogging about this further):

Adobe Premiere:
1. Imports FCP 7 stuff easily
2. Allows FCP 7 hot keys
3. Works with inexpensive NAS storage with no special (aka expensive) stuff
4. Listens to their customers and fixes other peoples’ bugs
5. Has a real integration engineer who actually helps with support issues and ropes in the right people.

If anyone had any doubts as to what the best choice was for a replacement to FCP 7, I’d like to suggest that even if there are things about Premiere you don’t like, if you speak up, they are likely to take you seriously and fix it. I feel pretty safe recommending Premiere when customers ask me about the best NLE to use these days.

Steve

[Todd Kopriva] “This is fixed in updates to Premiere Pro and Adobe Media Encoder and in the next version of After Effects:
https://blogs.adobe.com/aftereffects/2011/05/cant-create-quicktime-movie-lar...

———————————————————————————————————
Todd Kopriva, Adobe Systems Incorporated
After Effects quality engineering
After Effects team blog”

Steve Modica
CTO, Small Tree Communications