Researchers in South Korea have built a networking router that transmits data at record speeds from components found in most high-end desktop computers. A team from the Korea Advanced Institute of Science and Technology created the router, which transmits data at nearly 40 gigabytes per second--many times faster than the previous record for such a device. The techniques used by the researchers could lead to a number of breakthroughs, including the use of cheaper commodity chips, such as those made by Intel and Nvidia, in high-performance routers, in place of custom-made hardware. The software developed by the researchers could also serve as a testbed for novel networking protocols that might eventually replace the decades-old ones on which the Internet currently runs. http://www.technologyreview.com/communications/26096/?nlid=3423
On Sun, 22 Aug 2010 22:23:19 -1000, Michael Painter said:
Researchers in South Korea have built a networking router that transmits data at record speeds from components found in most high-end desktop computers http://www.technologyreview.com/communications/26096/?nlid=3423
Two great quotes from the article: "That isn't fast enough to take advantage of the full speed of a typical network card, which operates at 10 gigabytes per second." Anybody got a network of PCs that have cards that run at 10GBytes/sec? ;) For that matter, have enough 10Gbit network cards shipped that they are now considered "typical" (as in "more than 5%")? A Lamborghini costs about 10 times as much as a nice Camry. 10Gig cards are closer to 30-50 times as much as 1gig cards. Now, if Lambos aren't typical cars, are 10Gig cards typical? Just sayin'.... "Lash enough software routers together that run at 40 gigabytes per second, and you get what is essentially a single-terabit router. Using such a system, routers might some day run completely in software." Ahh.. but the lashing is the tricky part that costs the big bucks, as these guys will undoubtedly discover - life will get a lot more complicated once they saturate the first PCI backplane and need a second. Who wants to bet they'll end up re-inventing SGI's NUMAlink or similar interconnect? ;)
On Mon, 23 Aug 2010 05:59:43 -0400 Valdis.Kletnieks@vt.edu wrote:
On Sun, 22 Aug 2010 22:23:19 -1000, Michael Painter said:
Researchers in South Korea have built a networking router that transmits data at record speeds from components found in most high-end desktop computers http://www.technologyreview.com/communications/26096/?nlid=3423
Two great quotes from the article:
"That isn't fast enough to take advantage of the full speed of a typical network card, which operates at 10 gigabytes per second."
Anybody got a network of PCs that have cards that run at 10GBytes/sec? ;)
I missed that, and that answers the "was it a GigaBytes verses Gigabits error" question. Nothing new here by the looks of it - people in this thread were getting those sorts of speeds a year ago out of PC hardware under Linux - http://lkml.org/lkml/2009/7/15/234 "I have achieved a collective throughput of 66.25 Gbit/s." "We've achieved 70 Gbps aggregate unidirectional TCP performance from one P6T6 based system to another."
Mark Smith wrote:
On Mon, 23 Aug 2010 05:59:43 -0400 Valdis.Kletnieks@vt.edu wrote:
I missed that, and that answers the "was it a GigaBytes verses Gigabits error" question. Nothing new here by the looks of it - people in this thread were getting those sorts of speeds a year ago out of PC hardware under Linux -
http://lkml.org/lkml/2009/7/15/234
"I have achieved a collective throughput of 66.25 Gbit/s."
"We've achieved 70 Gbps aggregate unidirectional TCP performance from one P6T6 based system to another."
Very nice, but doing this with 1514-byte packets is the low-hanging fruit. (9K packets? That's the fruit that falls off the tree and into your basket while you're napping :-).) The more interesting limit: how many 40-byte packets per second can you shovel into this system and still have all of them come out the other end? Jim Shankland
Date: Mon, 23 Aug 2010 06:27:00 -0700 From: Jim Shankland <nanog@shankland.org>
Mark Smith wrote:
On Mon, 23 Aug 2010 05:59:43 -0400 Valdis.Kletnieks@vt.edu wrote:
I missed that, and that answers the "was it a GigaBytes verses Gigabits error" question. Nothing new here by the looks of it - people in this thread were getting those sorts of speeds a year ago out of PC hardware under Linux -
http://lkml.org/lkml/2009/7/15/234
"I have achieved a collective throughput of 66.25 Gbit/s."
"We've achieved 70 Gbps aggregate unidirectional TCP performance from one P6T6 based system to another."
Very nice, but doing this with 1514-byte packets is the low-hanging fruit. (9K packets? That's the fruit that falls off the tree and into your basket while you're napping :-).) The more interesting limit: how many 40-byte packets per second can you shovel into this system and still have all of them come out the other end?
Seems reasonable, but in our testing of 100G Ethernet capable routers we found one that handled 8000 bytes just fine, but could only run 9000 byte packets at about 90G. Just a bit unexpected. Really, in this day and age, a chassis throughput of 100G is pretty trivial. When you start getting up to the Tbps range on a system using "standard components", then I'll be really interested. We do have a network of many end systems attached with 10Gbps Ethernet cards. I'm sure that we are not unique, though probably unusual. We are achieving stable disk to disk transfer rates of well over 3G between the US and Australia. I don't think that PacketShader would handle the load too well. -- R. Kevin Oberman, Network Engineer Energy Sciences Network (ESnet) Ernest O. Lawrence Berkeley National Laboratory (Berkeley Lab) E-mail: oberman@es.net Phone: +1 510 486-8634 Key fingerprint:059B 2DDF 031C 9BA3 14A4 EADA 927D EBB3 987B 3751
Really, in this day and age, a chassis throughput of 100G is pretty trivial. When you start getting up to the Tbps range on a system using "standard components", then I'll be really interested.
i suspect that a rule of thumb is that leading edge home appliances are one decimal digit behind leading edge routers and cost two decimal digits less. so it's a trade-off. which is why we get the big bucks. randy
On 8/23/10 2:59 AM, Valdis.Kletnieks@vt.edu wrote:
On Sun, 22 Aug 2010 22:23:19 -1000, Michael Painter said:
Researchers in South Korea have built a networking router that transmits data at record speeds from components found in most high-end desktop computers http://www.technologyreview.com/communications/26096/?nlid=3423
Two great quotes from the article:
"That isn't fast enough to take advantage of the full speed of a typical network card, which operates at 10 gigabytes per second."
Anybody got a network of PCs that have cards that run at 10GBytes/sec? ;)
I have a journalist who can keep track of signficant digits...
For that matter, have enough 10Gbit network cards shipped that they are now considered "typical" (as in "more than 5%")? A Lamborghini costs about 10 times as much as a nice Camry. 10Gig cards are closer to 30-50 times as much as 1gig cards. Now, if Lambos aren't typical cars, are 10Gig cards typical? Just sayin'....
10gig nics are becoming ubiquitous in datacenters. 10Gigabit on mainboard is pretty ubiquitous in in bladeservers and adds about $150 to the BOM of a 1u pizza box in volume (for copper).
"Lash enough software routers together that run at 40 gigabytes per second, and you get what is essentially a single-terabit router. Using such a system, routers might some day run completely in software."
Ahh.. but the lashing is the tricky part that costs the big bucks, as these guys will undoubtedly discover - life will get a lot more complicated once they saturate the first PCI backplane and need a second. Who wants to bet they'll end up re-inventing SGI's NUMAlink or similar interconnect? ;)
pci isn't a shared bus anymore it's a series of tubes... In any event, they don't have to, we have quick-path which 100Gb/s per-direction per path at 3.2ghz or pci-e 3.0 which is 8Gb/s per lane and comes with all the lovely logic you expect from a (non-ethernet) switch fabric. What it really comes down to is packets per watt or packets per dollar, if it's cheaper to do it this way then people will, if not BFD.
On 8/23/2010 1:17 PM, Joel Jaeggli wrote:
What it really comes down to is packets per watt or packets per dollar, if it's cheaper to do it this way then people will, if not BFD.
I disagree here. Core routing isn't purchased based on cost, it's purchased based on support. People have not adopted Vayetta, or Mikrotik or many of the other small routing platforms which are in fact MUCH cheaper than the bridge or the tree (cisco or juniper), and the reason is simply support. If my router breaks beyond my ability to fix it I have a certified engineer (of some value or other) at my site with parts to fix it within 4 hours. This is why people go with Cisco and Juniper. It's also a mechanism of CYA. Would we rather tell our boss that the company has responded and dropped the replacement part in the mail, or that a technician from the router supplier is on their way and will be here very shortly, and ooh, by the way, you did recommend redundant hardware when the piece that broke was purchased, and it was available at a discount. Andrew
On Aug 23, 2010, at 12:25 PM, Andrew Kirch wrote:
On 8/23/2010 1:17 PM, Joel Jaeggli wrote:
What it really comes down to is packets per watt or packets per dollar, if it's cheaper to do it this way then people will, if not BFD.
I disagree here. Core routing isn't purchased based on cost, it's purchased based on support. People have not adopted Vayetta, or Mikrotik or many of the other small routing platforms which are in fact MUCH cheaper than the bridge or the tree (cisco or juniper), and the reason is simply support.
I disagree. Core routing is about performance, and, the bridge and the tree simply outperform Vayetta and Mikrotik on more realistic small packet sizes when it comes to forwarding rate, interface density, and other issues. Outside the core, you might be right about it being a question of support.
If my router breaks beyond my ability to fix it I have a certified engineer (of some value or other) at my site with parts to fix it within 4 hours. This is why people go with Cisco and Juniper. It's also a mechanism of CYA. Would we rather tell our boss that the company has responded and dropped the replacement part in the mail, or that a technician from the router supplier is on their way and will be here very shortly, and ooh, by the way, you did recommend redundant hardware when the piece that broke was purchased, and it was available at a discount.
That doesn't help as much as you might hope. I've had situations where it tool (bridge or tree) several months to resolve a problem. I have a case open with one of those vendors now for a PMTU-D problem which has been ongoing for many months. Often, I get an update saying it's been escalated to engineering, several weeks go by and I get a request for information already provided. Owen
On 8/23/10 12:25 PM, Andrew Kirch wrote:
On 8/23/2010 1:17 PM, Joel Jaeggli wrote:
What it really comes down to is packets per watt or packets per dollar, if it's cheaper to do it this way then people will, if not BFD.
I disagree here. Core routing isn't purchased based on cost, it's purchased based on support. People have not adopted Vayetta, or Mikrotik or many of the other small routing platforms which are in fact MUCH cheaper than the bridge or the tree (cisco or juniper), and the reason is simply support.
Neither of those are in the running for .5-1Tb/s forwarding devices. stack up enough vyatta boxs to equal an mx960 or a t1600 and I think you'll get my point.
If my router breaks beyond my ability to fix it I have a certified engineer (of some value or other) at my site with parts to fix it within 4 hours. This is why people go with Cisco and Juniper. It's also a mechanism of CYA. Would we rather tell our boss that the company has responded and dropped the replacement part in the mail, or that a technician from the router supplier is on their way and will be here very shortly, and ooh, by the way, you did recommend redundant hardware when the piece that broke was purchased, and it was available at a discount.
Andrew
participants (9)
-
Andrew Kirch -
Jim Shankland -
Joel Jaeggli -
Kevin Oberman -
Mark Smith -
Michael Painter -
Owen DeLong -
Randy Bush -
Valdis.Kletnieks＠vt.edu