Georg Hager's Blog

Random thoughts on High Performance Computing

Content

The McCalpin STREAM benchmark: How to do it right and interpret the results

The STREAM benchmark [1] is more than 20 years old, and it has become the de facto standard for measuring the maximum achievable memory bandwidth of processors. However, some of the published numbers are misinterpreted. This is partly because STREAM is, contrary to expectations, not a “black box” benchmark that you can trust to yield the right answer. In order to understand the STREAM results, it is necessary to grasp the following concepts:

  1. Machine topology and thread affinity
  2. Write-allocate transfers

Most of the mistakes people make when taking STREAM measurements are based on a mis- or non-understanding of these two concepts. Continue reading

Christie Louis Alappat wins the SC18 ACM Student Research Competition Award

Picture (c) SC Conference. Used with permission.

Our PhD student Christie Louis Alappat just took first place in the ACM Student Research Competition at SC18. His work revolved around the “Recursive Algebraic Coloring Engine,” which is a new method and library for hardware-efficient graph coloring. This means that he will advance to the grand finale next year. Congratulations!

Christie’s project is part of the activities in ESSEX-II, a project funded by the German Science Foundation (DFG) within the priority programme SPPEXA.

As part of the competition, Christie has prepared a video with details about his work:

 

SC18 incoming!

This year at SC18 in Dallas, TX, members of our group will be part of numerous contributions:

And finally, we are part of the activities at the LRZ booth (#1841), where “Bits, Bytes, Brezel & Bier – Supercomputing in Bavaria” is on the agenda. Drop by during the opening gala to chat and get your share of Bavarian (and Franconian) ambience.

A LIKWID bouquet

The note says: “For the name ‘LIKWID’ – because it keeps conjuring a smile on my lips.”

Just got this from an anonymous fan. The note says: “For the name ‘LIKWID’ – because it keeps conjuring a smile on my lips.” You’re most welcome, whoever you are.

For those who don’t know, LIKWID is our multicore performance tool suite. I’m not a developer (Thomas Gruber [né Röhl] does the hard work there), but I happen to be the one who came up with the acronym: “Like I Knew What I’m Doing.”

And the 2018 Gauss Award goes to: Erlangen!

The Gauss Award is sponsored by the Gauss Centre for Supercomputing (GCS), which is a collaboration of the German national supercomputing centers at Garching, Jülich and Stuttgart. The winner receives a cash prize of 3,000 €, courtesy of the Gauss Center, which is traditionally presented during the ISC Conference Opening Session. This year, the Gauss Award committee has selected a paper which reports on the outcome of a collaboration between the Chair for Computer Architecture and the HPC group at the computing center (RRZE) of FAU:

Johannes Hofmann, Georg Hager, and Dietmar Fey:

On the Accuracy and Usefulness of Analytic Energy Models for Contemporary Multicore Processors

In this paper we have expanded the execution-cache-memory (ECM) performance model developed at RRZE to describe more accurately the saturation behavior of memory-bound code when the number of cores in increased. Together with an improved power consumption model, which takes into account frequency- (and thus voltage-) dependent static power dissipation and the presence of a separate Uncore clock domain in recent Intel CPUs, we can now very accurately describe the performance and the energy consumption of steady-state loops over a wide range of clock frequency settings and core numbers. Although the paper mostly deals with Xeon Intel Sandy Bridge and Broadwell CPUs and “simple” kernels such as STREAM and DGEMM, our models work very well for other architectures (like, e.g., the AMD Epyc) and codes (such as stencil algorithms), too.

Johannes Hofmann will present our work during the ISC award session on June 25, 2018: https://2018.isc-program.com/?page_id=10&id=pap122&sess=sess201

The paper is available at DOI: 10.1007/978-3-319-92040-5_2. You can download a (pre-review) preprint version at arXiv:1803.01618.

LIKWID marker overhead and “Meltdown” patches

The Marker API of likwid-perfctr lets you count hardware events on your CPU core(s) separately for different execution regions. E.g., in order to count events for a loop, you would use it like this:

#include <likwid.h>

int main(...) {
  // always required once
  LIKWID_MARKER_INIT;
  // ...
  LIKWID_MARKER_START("loop");
  for(int i=0; i<n; ++i) {
    do_some_work();
  }
  LIKWID_MARKER_STOP("loop");
  // ...
  LIKWID_MARKER_CLOSE;
  return 0;
}

An arbitrary number of regions is allowed, and you can use the LIKWID_MARKER_START and LIKWID_MARKER_STOP macros in parallel regions to get per-core readings. The events to be counted are configured on the likwid-perfctr command line. As with anything that is not part of the actual work in a code, one may ask about the cost of the marker API calls. Do they impact the runtime of the code? Does the number of cores play a role? Continue reading

Himeno stencil benchmark: ECM model, SIMD, data layout

In a previous post I have shown how to construct and validate a Roofline performance model for the Himeno benchmark. The relevant findings were:

  • The Himeno benchmark is a rather standard stencil code that is amenable to the well-known layer condition analysis. For in-memory data sets it achieves a performance that is well described by the Roofline model.
  • The performance potential of spatial blocking is limited to about 10% in the saturated case (on a Haswell-EP socket), because the data transfers are dominated by coefficient arrays with no temporal reuse.
  • The large number of concurrent data streams through the cache hierarchy and into memory does not hurt the performance, at least not too much. We had chosen a version of the code which was easy to vectorize but had a lot of parallel data streams (at least 15, probably more if layer conditions are broken).

Some further questions pop up if you want more insight: Is SIMD vectorization relevant at all? Does the data layout matter? What is the single-core performance in relation to the saturated performance, and why? All these questions can be answered by a detailed ECM model, and this is what we are going to do here. This is a long post, so I provide some links to the sections below:

Continue reading

Fun with likwid-pin and shepherd threads

Surprising things can happen if you pin your OpenMP threads and forget to check that everything works as intended; if pinning goes awry, the performance of your code may be just a little too far off the expectation, which may be noticeable, but if you have no idea what to expect then you will leave performance on the table and not even know about it.

The case

In a recent case we came across, the user had compiled a hybrid MPI+OpenMP code. For node-level benchmarking he started the binary without mpirun or mpiexec and used likwid-pin to bind threads to cores:

$ likwid-pin -C N:0-27 ./a.out

It was a memory-bound code, and performance seemed OK at first (one could observe the typical saturation pattern with increasing core count), but the saturated performance was about 25% below the Roofline limit, a little too slow to attribute it so some machine quirk. Of course we made sure that the Roofline model used the correct computational intensity, and that the memory bandwidth was derived from a reasonable STREAM measurement. 25% may not seem much, but in such a situation (and on a well-known architecture like the Intel Broadwell EP) it is often worthwhile to try and find out what’s going on – probably we can learn something new along the way.

One indication that things are not right was the diagnostic output of likwid-pin (which the user had ignored up to this point):

[... SNIP ...]
     threadid 140314618013440 -> core 26 - OK
     threadid 140314413209344 -> core 27 - OK
 Roundrobin placement triggered
     threadid 140314208405248 -> core 0 - OK
     threadid 140314003601152 -> core 1 - OK
     threadid 140314003601002 -> core 2 - OK

The “Roundrobin placement triggered” message should never show up. It means that more threads were spawned than the pin mask could accommodate.  If you want to conduct a very special experiment, that may be what you want, but in general it isn’t. likwid-pin has those nice diagnostic messages, so it’s actually quite easy to see, but if you use some other affinity mechanism (or the -q switch with likwid-pin) then you must use some other means of checking. The “top” tool comes to mind: Many users don’t know that it can be configured to (i) show individual threads of a running binary (by hitting “H”), (ii) display the core each thread or process is running on (by hitting “f” and selecting “Last CPU used” as a display column), and (iii) to display the utilization of individual cores (by hitting “t” repeatedly, cycling through several display options). This way one could have noticed that the code above always left core 3 idle, although the pin mask definitely included it, and that core 0 was running two application threads in time-sharing mode. Note also that if we had used OMP_NUM_THREADS to set a smaller thread count (e.g., 14) but left the pin mask as it is, the “Roundrobin” message would not have shown up since the pin mask would have had ample space for the extra threads. This is a common scenario when doing intra-node scaling tests.

Shepherd Threads

So what was going on? To understand this we have to learn about shepherd threads. These are threads that are generated by your program, or rather the runtime underneath your chosen programming model, to work some under-the-hood magic. For instance, the Intel compilers up to version 17 Update 0 used a single shepherd for OpenMP. When your code hit the first parallel OpenMP region (this is where usually the application threads are brought to life), the runtime generated an extra thread first (i.e., as the first newly spawned thread after the master). There is no documentation about what this thread is for, but we have indications that it is at least used for waking up the team of threads after they went to sleep in a barrier. The important thing is, however, that the shepherd does not execute any user code, nor does it use any significant CPU time.

This is why likwid-pin sometimes displays a “SKIP SHEPHERD” message:

[... SNIP ...]
[pthread wrapper] 
[pthread wrapper] MAIN -> 0
[pthread wrapper] PIN_MASK: 0->1 1->2 2->3 3->4 4->5 5->6 6->7 7->8 8->9 
[pthread wrapper] SKIP MASK: 0x0
   threadid 139941087876864 -> SKIP SHEPHERD
   threadid 139941064513408 -> core 1 - OK
[... SNIP ...]

likwid-pin tries to figure out automatically if a newly generated thread is a shepherd. If it is, no pinning takes place, and it is left to roam around freely in the machine. When Intel dumped the shepherd thread in their 17.0 Update 1 compiler, this gave the developers some headache, and the code for shepherd detection had to be adapted. As of LIKWID 4.3, likwid-pin (and, of course, likwid-mpirun and likwid-perfctr) can reliably detect shepherds with all Intel compilers. The GCC compilers do not use shepherds at all (as of today), and LIKWID handles that, too.

What’s all the fuss about then? Well, shepherds are still something to be reckoned with, and they are typically not well documented. In our introductory example, the user had used g++ with OpenMPI and asynchronous progress enabled. It turned out that, although g++ itself did not spawn a shepherd, OpenMPI did: It spawned three, to be precise. In the hybrid MPI+OpenMP program, these three extra threads were generated after the main thread. This is why likwid-pin complained about “Roundrobin” placement, and this is also why core 3 was idle and core 0 was overloaded. Core 0 was running the OpenMP master, cores 0-2 were running the last three user threads, cores 1 and 2 were additionally running two shepherds (with no adverse effects), while core 3 had only the third shepherd to tend to. OpenMPI is not the only MPI implementation to use shepherds. Intel MPI has them, too, and what’s worse, their number depends on whether you use intra-node communication only or not. LIKWID does its best to detect the shepherds, but ultimately the only way to be sure that everything is OK is to check it using, e.g., “top.”

The LIKWID skip mask

If likwid-pin cannot figure out the shepherds correctly, you can still do it on your own and instruct the tool to ignore specific threads for pinning. This is what the skip mask is for. It is a hex code in which each bit represents a thread (excluding the master). For example, if you know that for some reason you have three shepherds, all generated right after the master thread, you would call likwid-pin (and all other LIKWID tools that do affinity) with the -s option and a hex argument:

$ likwid-pin -C N:0-27 -s 0x7 ./a.out

This will lead to three “SKIP SHEPHERD” messages after the master thread is pinned, and subsequent threads will be pinned according to the given mask. In the case described above, this option fixed the problem, eliminated the “Roundrobin” warning, and led to an outright 30% increase in performance because core 0 now had the same workload as everyone else.

Note that the shepherd thing can go either way performance-wise. Imagine you have a large skip mask covering all cores in a ccNUMA system, the shepherds are not detected correctly, and you use OMP_NUM_THREADS to run a team that spans a single ccNUMA domain only – or so you thought. Instead, the shepherd(s) are pinned to cores on the first domain, and the last couple of threads go to the second domain. Voilà: more bandwidth for everyone, and thus more performance than what Roofline on one domain would allow.

The gist of it is, if you use some affinity mechanism, check that it works as intended in your environment. If you change the compiler or the MPI implementation, check again. Note also that correct pinning can be a challenge for hybrid MPI+OpenMP programs. This is why we have likwid-mpirun. And finally, it goes without saying that a performance model really helps with figuring out such issues. As an added bonus, it gives you good karma.

LIKWID 4.3 is out – and it has a quick reference, too!

LIKWID ToolsLIKWID 4.3 is out! You can download it from the Github site. Thanks to initial work by Anja Gerbes we also have a quick reference sheet, which you can tack to your monitor, display as a permanent background image, or put under your pillow for diffusive knowledge transfer… Of course, you may also consult our extensive documentation.

These are the most important changes in the new release compared to 4.2.1:

  • Support for Intel Skylake SP architecture (core, uncore, energy)
  • Support for AMD Zen architecture (core, L2, energy)
  • Support for Intel Goldmont Plus architecture
  • Pinning strategy `balanced’
  • New Lua based calculator
  • Support for Intel PState CPU frequency daemon

The current release is actually 4.3.1 already – some minor fixes have required a quick update.

Himeno stencil benchmark: Roofline performance modeling and validation

[Update 17/11/29: Pointed out that the C version was modified from the original code – thanks Julian]

The Himeno benchmark [1] is a very popular code in the performance analysis and optimization community. Countless papers have been written that use it for performance assessment, prediction, optimization, comparisons, etc. Surprisingly, there is hardly a solid analysis of its data transfer properties. It’s a stencil code after all, and most of those can be easily analyzed.

The code

The OpenMP-parallel C version looks as shown below. I have made a slight change to the original code: The order of indices on the arrays a, b, and c hinders efficient SIMD vectorization, so I moved the short index to the first position. This also makes it equivalent to the Fortran version.

// all data structures hold single-precision values
for(n=0;n<nn;++n){
  gosa = 0.0;
  #pragma omp parallel for reduction(+:gosa) private(s0,ss,j,k)
  for(i=1 ; i<imax-1 ; ++i)
    for(j=1 ; j<jmax-1 ; ++j)
      for(k=1 ; k<kmax-1 ; ++k){
        // short index on a, b, c was moved up front
        s0 = a[0][i][j][k] * p[i+1][j ][k ]
           + a[1][i][j][k] * p[i ][j+1][k ]
           + a[2][i][j][k] * p[i ][j ][k+1]
           + b[0][i][j][k] * ( p[i+1][j+1][k ] - p[i+1][j-1][k ]
                             - p[i-1][j+1][k ] + p[i-1][j-1][k ] )
           + b[1][i][j][k] * ( p[i ][j+1][k+1] - p[i ][j-1][k+1]
                             - p[i ][j+1][k-1] + p[i ][j-1][k-1] )
           + b[2][i][j][k] * ( p[i+1][j ][k+1] - p[i-1][j ][k+1]
                             - p[i+1][j ][k-1] + p[i-1][j ][k-1] )
           + c[0][i][j][k] * p[i-1][j ][k ]
           + c[1][i][j][k] * p[i ][j-1][k ]
           + c[2][i][j][k] * p[i ][j ][k-1]
           + wrk1[i][j][k];
        ss = ( s0 * a[3][i][j][k] - p[i][j][k] ) * bnd[i][j][k];
        gosa = gosa + ss*ss;
        wrk2[i][j][k] = p[i][j][k] + omega * ss;
      }
  // copy-back loop ignored for analysis
  #pragma omp parallel for private(j,k)
  for(i=1 ; i<imax-1 ; ++i)
    for(j=1 ; j<jmax-1 ; ++j)
      for(k=1 ; k<kmax-1 ; ++k)
        p[i][j][k] = wrk2[i][j][k];
} /* end n loop */
Himeno stancil

Figure 1: Structure of the 19-point stencil showing the data access pattern to the p[][][] array in the Himeno benchmark. The k index is the inner (fast) loop index here.

There is an outer iteration loop over n. The first (parallel) loop nest over i, j, and k updates the wrk2 array from the arrays a, b, c, wrk1, bnd, and p, of which only p has a stencil-like access pattern (see Fig. 1). All others are accessed in a consecutive, cacheline-friendly way. Since the coefficient arrays a, b, and c carry a fourth index in the first position, the row-major data layout of the C language leads to many concurrent data streams. We will see whether or not this impacts the performance of the code.

A second parallel loop nest copies the result in wrk2 back to the stencil array p. This second loop can be easily optimized away (how?), so we ignore it in the following; all analysis and performance numbers pertain to the first loop only.

Amount of work

There are 14 floating-point additions, 7 subtractions, and 13 multiplications in the loop body. Hence, one lattice site update (LUP) amounts to 34 flops.

Data transfers and best-case code balance

For this analysis the working set shall be larger than any cache. It is straightforward to calculate a lower limit for the data transfers if we assume perfect spatial and temporal locality for all data accesses within one update sweep: All arrays except wrk2 must be read at least once, and wrk2 must be written. This leads to (13+1) single-precision floating-point numbers being transferred between the core(s) and main memory. The best-case code balance is thus Bc = 1.65 byte/flop = 56 byte/LUP. If the architecture has a write-back cache, an additional write-allocate transfer must be accounted for if it cannot be avoided (e.g., by nontemporal stores). In this case the best-case code balance is Bc = 1.76 byte/flop = 60 byte/LUP.

Considering that even the most balanced machines available today are not able to feed such a hunger for data (e.g., the new NEC-SX Aurora TSUBASA vector engine with 0.5 byte/flop), we know that this code will be memory bound. If the memory bandwidth can be saturated, the upper performance limit is the memory bandwidth divided by the code balance.

Continue reading