BM Seer Unofficial thoughts from an anonymous Sun employee

« not a CEC post... | Main | Ultra-FAST Cryptogra... »

UltraSPARC T2 and its NIU that'll be good for you

Tuesday Oct 09, 2007

This summer we announced the UltraSPARC T2 chip, but one of the things we didn't talk about much was the US T2's NIU. So let's look at some of the delivered results.

By the by, you'll see a lot more on performance results on this blog today. Yep it's launch day. Now many of my colleagues are at CEC bellying up to the buffets and dropping their money at the tables, some of us are at home working to show you the latest

The UltraSPARC T2 10GbE has an integrated NIU (10GbE Network Interface Unit, the 10GbE is silent ) which provides better performance and reduces CPU overhead of network traffic when compared to servers that must use NICs (network interface cards). The UltraSPARC T2's NIU has much lower latency which reduces CPU overhead.

10GbE transmit, maximum throughput is 36% higher performance and CPU efficiency is 23% better
10GbE receive, maximum throughput is almost twice the performance, exceeding x8 bus bandwidth by 16%

UltraSPARC T2 with NIU has the following measured results TX: 14.6 Gb/s; RX 18.2 Gb/s. In contract the Atlas NIC has the following measured results TX: 10.7 Gb/s; RX 9.4 Gb/s.

All performance tests were run by Sun and of course used Solaris 10.

... but what about standard benchmarks, ny advice is either get this blog in your RSS or check back every hour as, "happy days are here again"

[7] Comments

Like this post?

Posted at 07:00AM Oct 09, 2007 in Benchmarks
Tags: cmt networking t2

Comments:

Might be nice to see some of the important details required to get a system to achieve those numbers. Stuff like number of connections/flows, packet size...

Posted by rick jones on October 09, 2007 at 09:07 AM PDT #

it is a blog on a busy day, I'll get to it at some point.

Since you seem to understand these things, how about sharing some of your data on the platforms you work with?

Posted by BM Seer on October 09, 2007 at 10:14 AM PDT #

Well, one of the platforms I've been playing with is a Sol10U4 1.2 GHz T2000 with a Neptune/Atlas (which is the 2x10G and which is the 4x1G anyway?) but I don't think the numbers would make you terribly happy since I cannot seem to get netperf to run across both ports at the same time when I initiate it from the T2000.

Posted by rick jones on October 10, 2007 at 03:12 PM PDT #

AFAIK, unmodified netperf will not work on two ports on any platform. The test is done using Sun developed tool that doesn't have this limitation.You can use iperf for two port transmit.

Two port tests are done with 145 connections, 256KB socket buffer size, 8192 bytes message.

For tuning, see
http://www.solarisinternals.com/wiki/index.php/Networks#Tuning_Sun_Dual_10GbE_on_T1000.2FT2000

Posted by Alan Chiu on October 11, 2007 at 11:36 AM PDT #

I should have been more clear - I was using mutliple, concurrent netperf tests, but TCP connections could only be initiated via one port. UDP could be sent via both ports, and TCP connections could be accepted via both ports, but not initiated. Attemping to initiate a TCP connection via one of the ports would simply sit there. Sol10U4 without or with the nxge data corruption patch installed.

Posted by rick jones on October 12, 2007 at 08:01 AM PDT #

Rick,
You said "but TCP connections could only be initiated via one port" and "Attemping to initiate a TCP connection via one of the ports would simply sit there. " so can you initiate TCP connections from one port or not?

What I would do is see if I can ping client in both direction for both ports, and take it from there.

Posted by Alan Chiu on October 12, 2007 at 09:22 AM PDT #

So, I can run netperf on remote systems and those netperfs can establish connections to netserver on the T2000 via either port of the add-on card. TCP_STREAM, TCP_MAERTS, TCP_RR etc.

When I run netperfs on the T2000, ie the T2000 is the active connector, the only ones which work are the ones going through only one of the two ports.

All the while ping works just fine. The two ports are configures with IP's in separate IP subnets, and are connexted back-to-back with the load generator(s).

We should probably take this offline. Within the mac.com domain I am known as perfgeek.

Posted by rick jones on October 12, 2007 at 10:25 AM PDT #