-
this is a GREAT article on Thunderbolt shared networks
-
Bob,
The Ars Technica article is flawed. Mr. van Beijnum measured disk performance instead of networking performance. See this thread for accurate measurements (Thunderbolt 1): https://forums.macrumors.com/showthread.php?t=1657957 . Thunderbolt 2 may be even faster—it would be nice if somebody measured it using two new MacBook Pros.
So, the breakdown is that IPoTB is really fast, but its current implementation does all routing in software, which brings performance down if more than two computers are involved. I believe it is actually possible to hardware-accelerate IPoTB by using native PCIe packet switching, but we don’t know if Apple will implement that.
-
Do these 10-second-long iPerf transfer tests really negate AT review? Just curious.
On a side note: reminds of the excitement of using FireWire for “high speed” networking back in 1997. Unlike TB, FireWire supports tree-like topology, isochronous transfers, can connect up to 63 devices… Future held promises of 3.2 and 6.4Gbs speeds with full backward compatibility. Cheap cables! FAAAAAAST!!!
Cross-OS support: even MS jumped on the bandwagon.
Didn’t pan out: security (same DMA vulnerabilities as TB) and other issues, and of course, Ethernet was already there, and wasn’t proprietary.
If, say, IpoTB absolutely kills IP over 10GbE on costs, deployment ease and speeds in some scenarios – isn’t it still a huge uphill battle to get it supported on multiple OSs and hardware platforms?
Until then, isn’t IPoTB just a cute little niche thingie, much like what TB is today?
-
Alex,
The AT review is incorrect by itself. Iljitsch has measured the speed of MBA’s internal drive, which is slower than TB1 speed. Of course, the iperf tests do negate the review, but you could see that it was wrong even without the tests. You can read the comment section for additional information. I am surprised that this review has not been retracted. The mistake is quite ironic, as Iljitsch “contributes articles about network protocols as well as Apple topics”. I mean, of all AT contributors, he should have known better.
I agree with you that the whole TB story reminds a lot of the FW story. However, I think that this time it might pan out differently. Mostly because TB is backed by Intel. But there are other differences.
TB2 (20Gb/s) is faster and cheaper than 10GbE (2-port TB2 controller costs just $10), FW800 (0.8Gb/s) was slower and cost about the same as 1GbE. TB’s max cable length is 100m (10m cables available now), FW’s was 4.5m. Finally, and this is the killer, 1GbE was built into every single Apple’s computer, while 10GbE is built into none. Since you’d have a better chance of finding an Ethernet cable lying around, you’d be using 1GbE to transfer files. Now, to use 10GbE networking you’ll have to use 2 TB cables, 2 external 10GbE TB interfaces, and a Cat 6 cable. Which is quite more complex and much more expensive than using a single TB cable.
Cross-platform support and dedicated switches will only happen if Apple publishes IPoTB as a standard. Being Apple, they might well prefer to keep IPoTB to themselves as a “competitive advantage”; in that case, it will certainly remain a niche technology for direct connections or very small workgroups.
P.S. Oh, and FW did find use as a networking technology in F-22/F-35 🙂
-
Sergei,
Appreciate the response. You’re probably right about the AT review. I am too lazy to re-read it, but remember the impression of a quick-n-dirty test that wasn’t to be taken too seriously.
I do have a few petty issues with some of your points though. 🙂
[Sergei Yakovlev] “Mostly because TB is backed by Intel.”
Not just “backed” but “developed”, and then given up for exclusive use to Apple. Didn’t make sense to me then, still doesn’t now. Awesome tech with a fantastic potential usurped by a single company hell-bent on squeezing every penny out of it – and now several years later there is a big “FAILURE” stamp all over it: pathetic adoption.
OTOH FW was developed and “backed” by far more companies than TB was/ is, notably Sony and TI besides others. Marrying it to DV and putting it as an interface of choice on first digital camcorders quickly drove up its adoption – which I don’t see happening with TB. Years after its introduction, its adoption is still close to nil.
[Sergei Yakovlev] “TB2 (20Gb/s) is faster and cheaper than 10GbE (2-port TB2 controller costs just $10)”
Perhaps it’s cheaper between two MBPs. Or even in a five-seat workgroup with TB cables running wild – perhaps. Anything beyond that, with wiring, switching and routing included – still cheaper? How about scalability, does it look as good as with Ethernet?
Don’t get me wrong, Ethernet over TB is way cool for quick-n-dirty shoestring shared storage setups. Maybe even fast and maybe even reliable. We won’t know for a while. It just doesn’t seem to have much future.
-
[Sergei Yakovlev] “FW800 (0.8Gb/s) was slower and cost about the same as 1GbE.”
I don’t think 1GbE was nearly as ubiquitous or cheap in 1997 as it is now – part of the reason IPoFW made sense.
-
What that review did show was the potential latency issues with the network packet mode. That is the hurdle TB must climb versus standard network topologies. That is also the main issue I have with the current expectations of TB2. The bandwidth is great however the latency can really limit the results and cause potential troubleshooting nightmares. This would make the TB2 setup far more complicated than a standard network. All of this really reminds me of the Token Rings and the arguments back then and that of course didn’t last.
Eric-ADK
Tech Manager
support@adkvideoediting.com -
[Alex Gerulaitis] “Not just “backed” but “developed”, and then given up for exclusive use to Apple.”
Sorry, I’ve used the wrong word. What I meant to say was since Intel has developed TB (née Light Peak), they must be interested in it succeeding in the market. The exclusive use was just for one year. I think Apple’s involvement actually furthered TB adoption rather than impeded it. The slow adoption is due to Intel’s tight control of TB licensing—they have even forced some TB products off the market.
[Alex Gerulaitis] “Years after its introduction, its adoption is still close to nil.”
Well, the adoption could be better, but it’s not all doom and gloom. Take a look at these products:
https://www.blackmagicdesign.com/products/blackmagicproductioncamera4k
https://www.blackmagicdesign.com/products/ultrastudiothunderbolt (TB2)
https://www.avid.com/US/products/Pro-Tools-HD-Native
https://apogeedigital.com/products/symphony-64-thunderbridge.php
https://www.promise.com/promotion_page/promotion_page.aspx?region=en-global&rsn=100 (TB2)
https://www.promise.com/promotion_page/promotion_page.aspx?region=en-global&rsn=103 (TB2)It looks like the professional market is supporting TB alright.
The problem is, the consumers don’t need TB that much, as USB3 will do for most situations. So the GPU card makers and motherboard vendors have been reluctant to add TB. Intel must have realized that this can eventually lead to TB dying, so now they are pushing TB add-ons with their “Thunderbolt Ready” program:
Hopefully, the GPU card makers start to integrate TB directly on their cards, or at least provide internal DP connectors to eliminate unsightly loopback cables.
[Alex Gerulaitis] “Perhaps it’s cheaper between two MBPs. Or even in a five-seat workgroup with TB cables running wild – perhaps. Anything beyond that, with wiring, switching and routing included – still cheaper? How about scalability, does it look as good as with Ethernet?”
I agree that right now IPoTB is only useful for simple scenarios.
Technically, it is possible to build larger networks by connecting multiple Mac Pros together, with each Mac Pro fanning out to iMacs and MacBook Pros (the six-device limitation only applies to a single TB chain). And it would be cheaper than 10GbE: $330 10m (or even $40 2m) TB cable per host vs. $30 0.5m TB cable + $1000 10GbE TB interface per host, plus the cost of 10GbE switch.
However: 1) the performance would be really bad, as the current IPoTB implementation is software-based (the performance degrades with each additional routing hop—you can see that in the tests I’ve linked to above); 2) the total network throughput would be limited by a single 20Gb/s connection.
The first problem can be solved if Apple hardware-accelerates IPoTB by using native PCIe packet switching. I believe this is possible on current hardware. The second problem can only be solved if Apple publishes the IPoTB standard so that other companies can produce an IPoTB switch with full bandwidth.
-
[Alex Gerulaitis] “I don’t think 1GbE was nearly as ubiquitous or cheap in 1997 as it is now – part of the reason IPoFW made sense.”
Right, it made sense for FW400. However, when FW800 was introduced in 2003, 1GbE was already shipping in PowerMacs and PowerBooks, so it didn’t make that much sense anymore.
Now, the situation is different. There are no Apple computers with 10GbE, but all of them have TB. While FW400 provided a 4x theoretical speed-up over Fast Ethernet, TB provides a 10x theoretical speed-up over Gigabit Ethernet. And 10GbE are not coming down in price anytime soon—the market adoption has been slower than everybody expected. So, IPoTB makes even more sense than IPoFW had back in the day.
But, of course, IPoTB is not directly comparable to Ethernet due to the issues I’ve mentioned above.
-
[EricBowen] “What that review did show was the potential latency issues with the network packet mode.”
Eric,
The review did not show the latency was in the IPoTB. It could be in the SMB protocol or in the file system (and it probably is, with buffers overflowing because of the slow disks). TB has better latency than 10GbE because it is simpler and the packets (frames) are smaller.
If people really insist on testing IPoTB by copying files, they should copy between two SSD RAIDs with average write performance exceeding 1250 MB/s. Of course, in that case, 10GbE should also be tested by copying files.
-
With a 2 device direct connect only without any other device chaining yes. Add other devices and other clients to the TB network chain, then no way it’s lower with the latency issues I am seeing. Also keep in mind 10Gbe has low latency profiles that 1Gbe does not. Many may not even be aware or using those. Oh and 10Gbe has definitely come down in price in the last year. The adapters alone are $350 to $550 and the switches are now under $1K.
Eric-ADK
Tech Manager
support@adkvideoediting.com
Reply to this Discussion! Login or Sign Up