Niagara: UltraSPARC's Viagra?

I like Sun's SPARC and UltraSPARC processors. There's something about them that feels inherently "right". But, let's be honest, their price/performance is lack-lustre to say the least. All right, it sucks. There, I've said it.

Their scalabilty is excellent, as the success of Sun's 100+ CPU SF25K behemoths illustrates. And that's important because sometimes, for certain workloads, straightline performance isn't everything. You wouldn't use a Formula 1 car in a rally, for example.

But many workloads are highly or even embaressingly parallel. For example, a web server. Each "hit" is potentially a different thread. Super fast CPUs like AMD's Opteron mask this by being really quick, in an inefficient manner. That is, they mask the fact that today's CPUs are *much* faster than the memory systems to which they're connected, and so even the fastest CPU spends a lot of time twiddling its electronic thumbs, waiting for memory (be it main RAM, or cache). The trouble is that while these fast CPUs are stalled, waiting for data or instructions to be fetched from memory, they're doing nothing. Well, nothing useful: they still eat up electricity. Today's multi-core chips (e.g., UltraSPARC-IV and the latest Opterons) suffer from the same problem, although the stalling is limited to each core (i.e., a stall in one core has no effect on the other).

While Intel was playing the GHz game, Sun's engineers recognised that a better way to faster performance--or more throughput--was to tackle the problem differently. Rather than endlessly ramping up the clock speed for smaller and smaller performance increases, they took a different approach: when the CPU stalls because it is waiting for memory, why not switch the CPU to a different thread? Sun's research concluded that a CPU spends about 75% of its time waiting for memory, so a CPU with four threads could conceivably be kept busy much more often: while three threads are waiting for memory, the fourth is running. (I guess an analogy would be the difference between using different processes and different threads. The former is more expensive to switch state (waiting for memory), while the latter is less.) To put it simply, a 4 GHz chip is busy only 25% of the time (the rest is spent waiting for memory), so it has an effective speed of 1 GHz.

The next logical step after having a single-core CPU that can run four threads is a multi-core CPU where each core has four threads. Enter Niagara, the code name for Sun's first implementation of its CMT (Chip Multi Threading) architecture. Niagara has eight cores, each of which runs four threads. Yep, 32 threads on a single CPU, 8 of which run simultaneously (one on each core).

As reported at El Reg, the first machines using Niagara--which will probably have a product name of "UltraSPARC-T1"--are already in Beta testing (no, alas, I don't have access to one). Here's what psrinfo says on these systems:

$ ./psrinfo -vp
The physical processor has 8 cores and 32 virtual processors
The core 0 has 4 virtual processors (0, 1, 2, 3)
The core 1 has 4 virtual processors (4, 5, 6, 7)
The core 2 has 4 virtual processors (8, 9, 10, 11)
The core 3 has 4 virtual processors (12, 13, 14, 15)
The core 4 has 4 virtual processors (16, 17, 18, 19)
The core 5 has 4 virtual processors (20, 21, 22, 23)
The core 6 has 4 virtual processors (24, 25, 26, 27)
The core 7 has 4 virtual processors (28, 29, 30, 31)
UltraSPARC-T1 (clock 1080 MHz)

So, we're looking at a ballpark clock speed of 1GHz. Doesn't sound very impressive, until you figure in the eight multi-threaded cores. In terms of throughput, these puppies are better than an 8 GHz CPU!

With all these threads (especially in multi-socket machines!), it's a good job that Solaris eats threads for lunch!