RE: Ecosystems cost of additional specs (RE: Please re ad - proposed WG termination)
Michael Krause <[email protected]> Wed, 07 Sep 2005 07:12:55 -0700
| Newsgroups | gmane.ietf.ipoib |
|---|---|
| Message-ID | <[email protected]> |
--===============1396799687== Content-Type: multipart/alternative; boundary="=====================_375353279==.ALT" --=====================_375353279==.ALT Content-Type: text/plain; charset="us-ascii"; format=flowed BTW, as a data point, testing with a PCI-X 266 10 GbE NIC hits link rate with using regular 1500B Ethernet frames and de facto techniques such as checksum off-load (link rate for Ethernet is higher than link rate for IB since one delivers a 10 Gbps payload rate and the other delivers a 10 Gbps raw rate which is typically 7-8 Gbps depending upon if the 2KB MTU is used something that is also larger than standard Ethernet's frame). This has been measured on Itanium platforms running Unix so people's mileage will vary but clearly there are implementations that are able to deliver performance without requiring jumbo / large MSS. It may take a bit more CPU but then no real server network workload runs requires link rate all of the time - statistically, there are bursts when applications complete various stages of the work. In any case, I sense the hum is to support the work but the reality is if one wants performance, the work isn't that compelling and there is that other value-add that IB brings called RDMA that was intended to be used to deliver high-performance workloads that many of us believe is the proper and more valuable technology to make happen for IB than IP connectivity. Existing IP over IB implementations measured on low-end Itanium running Linux is at 3-3.5 Gbps which is very respectable and meets nearly all workload needs. Heck, real world workloads like NFS do not even generate that much network bandwidth utilization so it is not clear where exactly the problem lies sans perhaps micro benchmarks which will be impossible for IB to beat Ethernet without jumping to wider links or higher signaling rates (the minimum required given the differences in the physical layers). Mike --=====================_375353279==.ALT Content-Type: text/html; charset="us-ascii" <html> <body> <font size=3><br> BTW, as a data point, testing with a PCI-X 266 10 GbE NIC <b>hits link rate</b> with using regular <b>1500B</b> Ethernet frames and de facto techniques such as checksum off-load <i>(link rate for Ethernet is higher than link rate for IB since one delivers a 10 Gbps payload rate and the other delivers a 10 Gbps raw rate which is typically 7-8 Gbps depending upon if the 2KB MTU is used something that is also larger than standard Ethernet's frame).</i> This has been measured on Itanium platforms running Unix so people's mileage will vary but clearly there are implementations that are able to deliver performance without requiring jumbo / large MSS. It may take a bit more CPU but then no real server network workload runs requires link rate all of the time - statistically, there are bursts when applications complete various stages of the work. <br><br> In any case, I sense the hum is to support the work but the reality is if one wants performance, the work isn't that compelling and there is that other value-add that IB brings called RDMA that was intended to be used to deliver high-performance workloads that many of us believe is the proper and more valuable technology to make happen for IB than IP connectivity. Existing IP over IB implementations measured on low-end Itanium running Linux is at 3-3.5 Gbps which is very respectable and meets nearly all workload needs. Heck, real world workloads like NFS do not even generate that much network bandwidth utilization so it is not clear where exactly the problem lies sans perhaps micro benchmarks which will be impossible for IB to beat Ethernet without jumping to wider links or higher signaling rates (the minimum required given the differences in the physical layers).<br><br> Mike</font></body> </html> --=====================_375353279==.ALT-- --===============1396799687== Content-Type: text/plain; charset="us-ascii" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit Content-Disposition: inline _______________________________________________ IPoverIB mailing list [email protected] https://www1.ietf.org/mailman/listinfo/ipoverib --===============1396799687==--