I have no idea what SUN’s talking about here, really. Supercomputing Online wasn’t helpful either. SUN introduced a 10GE NIC called to be “multi-threaded”, whatever the hell that’s supposed to mean; it’s their “first network interface specifically designed to accelerate multi-threaded application performance by optimizing I/O throughput within environments that utilize parallel threads“.
Rest. Silence. Taa-daa! So, dissecting this sentence their telling us that their NIC runs better with – yeah, what? POSIX threads? Java? I don’t get their point. NIC-drivers run in kernel-context and offer their services to every application – multi-threaded or not – with the same friggin’ performance. I don’t see why multi-threaded application should get more bandwidth than any other application.
As usal, Supercomputing Online didn’t bother to give a link to SUN’s website so I had to investigate what the frack is that all about. SUN’s website had that big ad just on the first page, “Unleashing 10GB Everywhere”. So let’s dive deeper into the topic, after ranting so much. (Too much. I had quite a sorrow day, please accept my apologis for beeing a sarcastic smartass)
The whole thing is about their “Complete chip multithreading environment (CMT)” – claiming that they offer full 10GE line-speed from application-level and beyond. A bold claim! Let’s read on. Later they say:
“Metaphorically, you can view the Sun Multithreaded 10 GigE Networking Technology as an ‘impedance matching device’ that extends the thread parallelism from the OS, through the processor, and all the way to the network wire,” says Shimon Muller, Sun distinguished engineer.”
“Impedance matching device-what-the-hell”? This ain’t even sales-speek to me, that’s more like very bad geek-speek. But then they come to the nitty-gritty:
“The Solaris OS has been a multithreaded system for years, and the latest UltraSPARC processors offer powerful chip multithreading,” says Sunay Tripathi, Sun distinguished engineer. “But what has been missing is extending multithreading into the I/O environment and networking space.”
Ariel Hendel, Sun distinguished engineer, explains that “The hardware /software interface up until now was based on first queuing packets into the system and then distributing and processing them. But with this new 10Gb/s technology, we move to a distribute then queue model, and that makes all the difference when you want to scale, apply policy, and so on.”
OK, now it’s getting interesting. They’re putting they queuing into the silicon. Reminds me of Tagged Command Queuing for NICs. That might give a couple of advantages, considering traffic-policies. However I’m waiting for benchmarks.
Is it interesting? Certainly, yes, especially that they annouce that the new CPUs will have two 10GE ports. If their CMT can really take advantage of offloading some stack’s routines to the NIC’s chip this might reduce latency from my naive point of view. But I can’t see Supercomputing Online’s intitial statement about “optimizing I/O throughput“. Throughput ain’t latency.
But I’m still curious! If you get a hand on this beauty: Drop me a line about your experiences.