# Preview for SC15 tutorial on “Node-Level Performance Engineering” now available

SC15 solicits video previews of accepted tutorials for the first time this year. So watch the commercial for our SC15 full-day tutorial “Node-Level Performance Engineering“!

Kudos to Jörn Rüggeberg from the RRZE multimedia center for putting together this great piece of art.

# Fooling the masses – Stunt 10: Always emphasize the “interesting” part of your work!

(See the prelude for some general information on what this is all about)

Have you ever thought about how to improve the aerodynamic properties of bulldozers? Yes? Then this stunt is for you.

Figure 1: Always focus on the interesting stuff!

Code optimization is quite rewarding at times. You put all your effort into mutilating a chunk of code beyond all recognition, and in the end you get a 50% speedup. Often enough, this work requires a great deal of  expertise at the interface between hardware and software. And you certainly want to present your success to your peers so they understand how advanced you are.

If you do such a thing, don’t bother picking the low-hanging fruits. When 95% of the run time goes into a boring loop nest which multiplies a dense matrix with a vector, there is not much you can do beyond the standard textbook stuff – or use a library altogether. However, if the remaining 5% consist of a convoluted communication scheme, you’ve found your target. The fact that the possible gain is at most 5% shouldn’t stop you, because it’s all a matter of presentation; see Fig. 1: 3D bars and a proper choice of scales on the axes guide the reader’s eye and help you drive home your point. It’s like presenting the optimized drag coefficient of your aerodynamically improved bulldozer on high gloss paper, and the boring technical data in small black print on a dark blue background, on the back page.

Interestingly, if the optimization target is communication, or synchronization, or any other non-computational overhead, this stunt can be made to work even better: When you strongly scale to a pointless number of workers, bringing down such overheads will give you arbitrary speedups, far beyond the factor of two that is theoretically possible if you restrict yourself to a sensible minimum parallel efficiency of 50%. Coming back to the heavy gear analogy, if you accelerate your bulldozer to 200 mph, the drag coefficient does become essential, and you may feel vindicated. It won’t solve any real-world problem, but hey, this is research after all.

Thanks to Gerhard Wellein for discovering this stunt in a paper. Many papers, actually.

# A case for the non-temporal store

Non-temporal stores (also called “streaming stores”) were added to x86 processors with SSE2 (i.e. when the first Pentium IV called “Willamette” appeared in 2000). There was also some support in MMX before, but nobody should be required to use MMX any more. In contrast to standard store instructions which transfer data from registers to memory, NT stores do not require a prior cache line read for ownership (RFO) but write to memory “directly” (well not quite – but to first order it’s true). Beware the usual explanations for RFO that Google finds for you – they almost all describe RFO as something inherently connected to cache coherency in multi-processor systems, but RFO is also required on a single core, whenever there is a cache that is organized in lines.

What are NT stores good for? Again, most references cite them as a means to avoid cache pollution in cases where stored data will not be re-used soon. That’s true to some extent; every word that you don’t store in your cache means more room for other, more frequently used data. But that’s only one half the story. Imagine you have a code whose working set fits into the outer-level cache and which produces data to be stored in memory. Using NT stores for such a code will probably slow it down because, depending on the LD/ST balance, performance is then limited by the memory interface. The true benefit of NT stores can be seen with a truly memory-bound code which is not dominated by loads, like, e.g., the STREAM COPY benchmark, relaxation methods, the lattice-Boltzmann algorithm, or anything that can be formulated in terms of stride-one store streams: By saving the RFO, the pressure on the memory subsystem is reduced. See, e.g., my talk at the KONWIHR results and review workshop in December 2007 at LRZ where the performance boost through NT stores was demonstrated using a 2D Jacobi (5-point stencil) code.

The most important non-temporal store instruction is movntpd, which writes the contents of a full 16-byte SSE register (xmm0…xmm15) to memory. The memory address has to be a multiple of 16, else an exception will be generated. Also, this instruction exists in a “packed” variant only, i.e. there is no movntsd that would only store the lower 8 bytes of the register (AMD has included it, though, into their SSE4a extensions which are implemented with the K10 CPUs – read a nice writeup by Rahul Chaturvedi from AMD for more information). So, if you are stuck with Intel processors or a compiler which does not know about movntsd on K10, loops which use NT stores must be vectorizable and the store stream(s) must be 16-byte aligned.

Or so I thought, for a long time. This assumption was backed, of course, by stupid compilers who blatantly refused to use NT stores if they weren’t 100% sure at compile time whether the store stream was actually aligned. Never heard of loop peeling, eh? Thank goodness great Intel gave us the vector aligned pragma, which you can put in front of a loop in order to tell the compiler that the streams appearing inside the loop were 16-byte aligned! But that always refers to all streams, including reads, and if you somehow forget to pay proper attention, you’re immediately punished by an exception again. To make a long story short, you never really know what’s going on inside the compiler’s twisted little brain, and the diagnostics don’t help either (“LOOP WAS VECTORIZED”, yes, thank you very much).

In hope for some future improvement (unaligned NT store? Or *gasp* maybe even a scalar one?) I looked into Intel’s AVX extensions for SSE, to be implemented in the Sandy Bridge processor, which is to follow Westmere in 2010. That would be a “Streaming SIMD Extension Extension”, right? Whatever, there is neither a scalar nor an unaligned packed NT store in there. But I stumbled across something that had been there since SSE2: The beautiful maskmovdqu instruction. From the manual:

MASKMOVDQU xmm1,xmm2 – Store Selected Bytes of Double Quadword with NT Hint

Stores selected bytes from the source operand (first operand) into an 128-bit memory location. The mask operand (second operand) selects which bytes from the source operand are written to memory. The source and mask operands are XMM registers. The location of the first byte of the memory location is specified by DI/EDI/RDI and DS registers. The memory location does not need to be aligned on a natural boundary. […]

Way cool. Although the fixed address register is somewhat inflexible, the instruction is even better than pure scalar or pure unaligned NT store – it can store any part of an XMM register, down to the byte level, and with no alignment restrictions. And there’s a C/C++ compiler intrinsic, too. That has made it easy to implement our beloved vector triad in explicit SSE. These are two possible variants:

 Scalar (unaligned) Packed unaligned  __m128d xmm1,xmm2,xmm3,xmm4; __m128i mask_low=_mm_set_epi32(0,0,-1,-1); #pragma omp parallel private(j) { if(size > 1000000) { for(j=0; j(xmm4), mask_low, (char*)(a+i)); } } } else { // same with standard store }  __m128d xmm1,xmm2,xmm3,xmm4; __m128i mask_all=_mm_set_epi32(-1,-1,-1,-1); #pragma omp parallel private(j) { if(size > 1000000) { for(j=0; j(xmm4), mask_all, (char*)(a+i)); } } } else { // same with standard store }

The left one is an example for a purely scalar code which employs NT stores. The right one shows how a vectorized code can be endowed with NT stores if 16-byte alignment cannot be enforced. The compiler does neither by itself. Of course, NT stores only make sense if the arrays are large. I have set the lower limit to N=1000000, which is largish compared to the cache sizes but neatly shows the effect of NT vs. standard store:

The figure includes standard scalar versions without NT stores as well. Obviously, the maskmovdqu instruction does the job. There is hardly any performance loss in the serial case, and none at all with four threads. I have omitted any comment on in-cache behavior of the different variants; this is a separate story (and one which becomes more interesting once Nehalem is out!).

If you want to know more about using SSE intrinsics in C/C++ code, the Intel C++ Compiler for Linux Intrinsics Reference is for you.

# Array summation benchmark

A question came up on the OpenMP mailing list today concerning scalability of simple array summation on an Opteron processor. I have done some tests with the following code, using the Intel C++ compiler version 9.1:

#pragma omp parallel for private(j) reduction(+: sum)
#pragma vector always
for (j = 0; j < N; j++){
sum += array2[j];
}

There is a loop around that to ensure that for small sizes we actually see the cache effect. Here is the result:

The number of threads (1T, 2T,…) is indicated. In case of the Opteron system, this was a 2-socket dual-core 2GHz box and the 2-thread data was correspondingly measured on one (1S) or two (2S) sockets, respectively. Proper NUMA placement was implemented. The “Conroe” system is my standard Core2 workstation.

Data on purely serial runs (no -openmp) is shown for reference. In contrast to low-level benchmarks like the stream or vector triads which have more read streams and at least one write stream, there seems to be a lot of “headroom” for the second thread even for large N on an Opteron socket.