P.S. Even at 1Gb we experienced problems in the Sun Streaming System moving blocks to storage over TCP/IP. The default TCP timers react outrageously slowly to packet drop events. So we'd get lots of 1MB blocks transmitted at near-wire speed, then one that would just blow up and take an order of magnitude or two longer latency.
We hacked the TCP timers.
But, I can understand why this sort of behavior would be unacceptable in a SAN.
Dangit, I kept trying to not use "data" that way but it looks like I slipped up twice. "interconnect" is even worse, it's really about WAN and inter-server traffic versus disk-bound traffic.
I fixed the original post to use "LAN/WAN" in one case (yay alphabet soup buzzwords!) and "IP interconnection" in the other.
The IP/storage dichotomy is a little better than data/storage but still inaccurate as there are people using iSCSI or other IP-based storage protocols today.
How do you...freelikebeerFebruary 15 2008, 15:37:05 UTC
Although advances in Ethernet may come slowly, Infiniband lacks a foothold in today's data centers, said Dave Zabrowski, chief executive of Neterion Inc. (Cupertino, Calif.), a startup making 10Gbit/s Ethernet adapters.
... even make the decision to start up a company to make 10GB ethernet adapters. It seems like you are building a commodity product that has a high capital investment cost, and a tight, competitive market against big industrial focused companies.
Re: How do you...markgritterFebruary 15 2008, 16:34:31 UTC
I wonder that, too... it's gotta be a time-to-market play, but until the big guys get into the market you don't have any switches to connect to. On the other hand, the big network adapter vendors started as startups too. :)
My advisor's current startup company makes a 10Gb switch (and the rumor is that Andy put some money in) so they evidently think that there is some room in the 10Gb market.
I think it's a reasonable gamble to be on the leading edge of a commodity product, and hope that either you become established in the high-end market before the bigger competitors, or get bought by one whose own development process is falling behind. There are some companies that try to make this work in the x64-box market, which has many of the same issues.
One issue with converting to either/or is that on the hardware side the protocols operate at different frequencies. For any clocking in the systems, this now means that there need to be buffers that are doing the correct conversion or that all the hardware on one side or the other needs to change speed.
From my work with Infiniband vs. Ethernet, the folks building for Infiniband were very locked into the planned speed increases (every incremental speed is done in steps of 2 versus Ethernet speed done in steps of 10). I don't know enough about the hardware on the SAN side to say definitively why these speeds are used, but I suspect it will be a long and difficult conversion and is most likely to involve some sort of translation due to the native speeds of the hardware involved and backwards compatibility.
Comments 8
We hacked the TCP timers.
But, I can understand why this sort of behavior would be unacceptable in a SAN.
Reply
(The comment has been removed)
Reply
(The comment has been removed)
The IP/storage dichotomy is a little better than data/storage but still inaccurate as there are people using iSCSI or other IP-based storage protocols today.
Reply
... even make the decision to start up a company to make 10GB ethernet adapters. It seems like you are building a commodity product that has a high capital investment cost, and a tight, competitive market against big industrial focused companies.
Reply
My advisor's current startup company makes a 10Gb switch (and the rumor is that Andy put some money in) so they evidently think that there is some room in the 10Gb market.
I think it's a reasonable gamble to be on the leading edge of a commodity product, and hope that either you become established in the high-end market before the bigger competitors, or get bought by one whose own development process is falling behind. There are some companies that try to make this work in the x64-box market, which has many of the same issues.
Reply
From my work with Infiniband vs. Ethernet, the folks building for Infiniband were very locked into the planned speed increases (every incremental speed is done in steps of 2 versus Ethernet speed done in steps of 10). I don't know enough about the hardware on the SAN side to say definitively why these speeds are used, but I suspect it will be a long and difficult conversion and is most likely to involve some sort of translation due to the native speeds of the hardware involved and backwards compatibility.
Reply
Reply
Reply
Leave a comment