The Death of Infiniband
I've always wanted to predict the death of something.
Today is my lucky day - I predict the death of Infiniband.
Yeah, sure. I'm not exactly going out on a limb here. Consolidation onto Ethernet has been an ongoing saga in networking as one LAN technology after another dies and is replaced with Ethernet. FDDI, ATM, Token Ring, and even early AppleTalk. The list goes on...
However, there have been some things that even Ethernet hasn't mastered yet and low latency is one of them. The market for low latency networking has continued amongst the High Performance Computing (HPC) crowd where researchers who can't brute force their way through gigabit Ethernet or afford big SMP machines end up with either Infiniband or Myrinet.
The problem with the HPC space, however, is that it isn't very big to start with and the growth rate just isn't there. As is traditionally the case, the real growth opportunity lies in the enterprise. The challenge with the enterprise for high speed interconnects is that they aren't ready to blow off their investment in Ethernet. As a result, Myrinet has already announced a 10G Ethernet product. They've made their position pretty clear: we aren't going to beat'm, so we'll join'm. They'll still offer their uber-low-latency products, but administrators will have a choice.
Mellanox, the Infiniband poster child, may be pushing their Infiniband story in the press, however they recently backed a 40G Ethernet proposal. I'm not a genius, but if that don't hint to a roadmap, I don't know what does.
In the end, the only player that will be left for the foreseeable future will be Fibre Channel (FC). This isn't because FC has the magic touch that will make it invincible to Ethernet, but it is because FC has a significant installed base that isn't going to toss their investment so quickly. But... If FC over Ethernet (FCoE) takes off as expected, it is only a matter of time before straight up FC goes the way of SNA.
Does all this death create a significant opportunity for low latency Ethernet? That remains to be seen. To date, I have only seen low latency claims achieved through RDMA and I don't buy that applications are going to change to accommodate the protocol shift. Does intra-datacenter latency really matter then? Will people put money behind reduced latency Ethernet that doesn't require changes?
Good questions. I believe that the answer is yes. But that's for another blog entry...
Today is my lucky day - I predict the death of Infiniband.
Yeah, sure. I'm not exactly going out on a limb here. Consolidation onto Ethernet has been an ongoing saga in networking as one LAN technology after another dies and is replaced with Ethernet. FDDI, ATM, Token Ring, and even early AppleTalk. The list goes on...
However, there have been some things that even Ethernet hasn't mastered yet and low latency is one of them. The market for low latency networking has continued amongst the High Performance Computing (HPC) crowd where researchers who can't brute force their way through gigabit Ethernet or afford big SMP machines end up with either Infiniband or Myrinet.
The problem with the HPC space, however, is that it isn't very big to start with and the growth rate just isn't there. As is traditionally the case, the real growth opportunity lies in the enterprise. The challenge with the enterprise for high speed interconnects is that they aren't ready to blow off their investment in Ethernet. As a result, Myrinet has already announced a 10G Ethernet product. They've made their position pretty clear: we aren't going to beat'm, so we'll join'm. They'll still offer their uber-low-latency products, but administrators will have a choice.
Mellanox, the Infiniband poster child, may be pushing their Infiniband story in the press, however they recently backed a 40G Ethernet proposal. I'm not a genius, but if that don't hint to a roadmap, I don't know what does.
In the end, the only player that will be left for the foreseeable future will be Fibre Channel (FC). This isn't because FC has the magic touch that will make it invincible to Ethernet, but it is because FC has a significant installed base that isn't going to toss their investment so quickly. But... If FC over Ethernet (FCoE) takes off as expected, it is only a matter of time before straight up FC goes the way of SNA.
Does all this death create a significant opportunity for low latency Ethernet? That remains to be seen. To date, I have only seen low latency claims achieved through RDMA and I don't buy that applications are going to change to accommodate the protocol shift. Does intra-datacenter latency really matter then? Will people put money behind reduced latency Ethernet that doesn't require changes?
Good questions. I believe that the answer is yes. But that's for another blog entry...
posted by Steve Shah at 9:31 AM
0 Comments:
Post a Comment
<< Home