Vampir 9.3

A Use Case

This chapter explains by example how Vampir can be used to discover performance problems in your code and how to correct them.

Introduction

In many cases the Vampir suite has been successfully applied to identify performance bottlenecks and assist their correction. To show in which ways the provided toolset can be used to find performance problems in program code, one optimization process is illustrated in this chapter. The following example is a three-part optimization of a weather forecast model including simulation of cloud microphysics. Every run of the code has been performed on 100 cores with manual function instrumentation, MPI communication instrumentation, and recording of the number of L2 cache misses.

Master Timeline and Function Summary showing an overview of the program run
Master Timeline and Function Summary showing an overview of the program run

Getting a grasp of the program's overall behavior is a reasonable first step. In Figure LINK Vampir has been set up to provide such a high-level overview of the model's code. This layout can be achieved through two simple manipulations. Set up the Master Timeline to adjust the process bar height to fit the chart height. All 100 processes are now arranged into one view. Likewise, change the event category in the Function Summary to show function groups. This way the many functions are condensed into fewer function groups.

One run of the instrumented program took 290 seconds to finish. The first half of the trace (Figure LINK A) is the initialization part. Processes get started and synced, input is read and distributed among these processes. The preparation of the cloud microphysics (function group: MP) is done here as well.

The second half is the iteration part, where the actual weather forecasting takes place. In a normal weather simulation this part would be much larger. But in order to keep the recorded trace data and the overhead introduced by tracing as small as possible only a few iterations have been recorded. This is sufficient since they are all doing the same work anyway. Therefore the simulation has been configured to only forecast the weather 20 seconds into the future. The iteration part consists of two "large" iterations (Figure LINK B and C), each calculating 10 seconds of forecast. Each of these in turn is partitioned into several "smaller" iterations.

For our observations we focus on only two of these small, inner iterations, since this is the part of the program where most of the time is spent. The initialization work does not increase with a higher forecast duration and would only take a relatively small amount of time in a real world run. The constant part at the beginning of each large iteration takes less than a tenth of the whole iteration time. Therefore, by far the most time is spent in the small iterations. Thus they are the most promising candidates for optimization.

All screenshots starting with Figure LINK are in a before-and-after fashion to point out what changed by applying the specific improvements.

Identified Problems and Solutions

Computational Imbalance

A varying size of work packages (thus varying processing time of this work) means waiting time in subsequent synchronization routines. This section points out two easy ways to recognize this problem.

Before Tuning: Master Timeline and Function Summary identifying MICROPHYSICS (purple color) as predominant and unbalanced
Before Tuning: Master Timeline and Function Summary identifying MICROPHYSICS (purple color) as predominant and unbalanced

After Tuning: Timeline and Function Summary showing an improvement in communication behavior
After Tuning: Timeline and Function Summary showing an improvement in communication behavior

Problem

As can be seen in Figure LINK each occurrence of the MICROPHYSICS-routine (purple color) starts at the same time on all processes inside one iteration, but takes between 1.7 and 1.3 seconds to finish. This imbalance leads to idle time in subsequent synchronization calls on the processes 1 to 4, because they have to wait for process 0 to finish its work (marked parts in Figure LINK). This is wasted time which could be used for computational work, if all MICROPHYSICS-calls would have the same duration. Another hint at this overhead in synchronization is the fact that the MPI receive routine uses 17.6% of the time of one iteration (Function Summary in Figure LINK).

Solution

To even out this asymmetry the code which determines the size of the work packages for each process had to be changed. To achieve the desired effect an improved version of the domain decomposition has been implemented. Figure LINK shows that all occurrences of the MICROPHYSICS-routine are vertically aligned, thus balanced. Additionally the MPI receive routine calls are now clearly smaller than before. Comparing the Function Summary of Figure LINK and Figure LINK shows that the relative time spent in MPI receive has been decreased, and in turn the time spent inside MICROPHYSICS has been increased greatly. This means that we now spend more time computing and less time communicating, which is exactly what we want.

Serial Optimization

Inlining of frequently called functions and elimination of invariant calculations inside loops are two ways to improve the serial performance. This section shows how to detect candidate functions for serial optimization and suggests measures to speed them up.

Problem

All performance charts in Vampir show information of the time span currently selected in the timeline. Thus the most time-intensive routine of one iteration can be determined by zooming into one or more iterations and having a look at the Function Summary. The function with the largest bar takes up the most time. In this example (Figure LINK) the MICROPHYSICS-routine can be identified as the most costly part of an iteration. Therefore it is a good candidate for gaining speedup through serial optimization techniques.

Solution

In order to get a fine-grained view of the MICROPHYSICS-routine's inner workings we had to trace the program using full function instrumentation. Only then it was possible to inspect and measure subroutines and subsubroutines of MICROPHYSICS. This way the most time consuming subroutines have been spotted, and could be analyzed for optimization potential.

The review showed that there were a couple of small functions which were called a lot. So we simply inlined them. With Vampir you can determine how often a functions is called by changing the metric of the Function Summary to the number of invocations.

The second inefficiency we discovered had been invariant calculations being done inside loops. So we just moved them in front of the respective loops.

Figure LINK sums up the tuning of the computational imbalance and the serial optimization. In the timeline you can see that the duration of the MICROPHYSICS-routine is now equal among all processes. Through serial optimization the duration has been decreased from about 1.5 to 1.0 second. A decrease in duration of about 33% is quite good given the simplicity of the changes done.

High Cache Miss Rate

Before Tuning: Counter Data Timeline revealing a high amount of L2 cache misses inside the CLIPPING-routine (light blue)
Before Tuning: Counter Data Timeline revealing a high amount of L2 cache misses inside the CLIPPING-routine (light blue)

After Tuning: Visible improvement of the cache usage
After Tuning: Visible improvement of the cache usage

The latency gap between cache and main memory is about a factor of 8. Therefore optimizing for cache usage is crucial for performance. If you don't access your data in a linear fashion as the cache expects, so called cache misses occur and the specific instructions have to suspend execution until the requested data arrives from main memory. A high cache miss rate therefore indicates that performance might be improved through reordering of the memory access pattern to match the cache layout of the platform.

Problem

As can be seen in the Counter Data Timeline (Figure LINK) the CLIPPING-routine (light blue) causes a high amount of L2 cache misses. Also its duration is long enough to make it a candidate for inspection. What caused these inefficiencies in cache usage were nested loops, which accessed data in a very random, non-linear fashion. Data access can only profit from cache if subsequent read calls access data in the vicinity of the previously accessed data.

Solution

After reordering the nested loops to match the memory order, the tuned version of the CLIPPING-routine now needs only a fraction of the original time. (Figure LINK)

Conclusion

By using the Vampir toolkit, three problems have been identified. As a consequence of addressing each problem, the duration of one iteration has been decreased from 3.5 seconds to 2.0 seconds.

Overview showing a significant overall improvement
Overview showing a significant overall improvement

As is shown by the Ruler, see Section LINK, in Figure LINK two large iterations now take 84 seconds to finish. Whereas at first (Figure LINK) it took roughly 140 seconds, making a total speed gain of 40%.

This huge improvement has been achieved by using the insight into the program's runtime behavior, provided by the Vampir toolkit, to optimize the inefficient parts of the code.