Homework 2 - CIS371 Spring 2008

Instructor:Prof. Milo Martin
Due:3pm, Monday, April 14th (at the start of class)


In this assignment you will explore the effectiveness of branch direction prediction (taken vs not taken) on an actual program. We've used Pin, a binary instrumentation tool, to generate a trace of branches and their outcomes. Your task will be use this representative trace to evaluate the effectiveness of a few simple branch prediction schemes. To do this, you'll write a program that reads in the trace and simulates different branch prediction schemes and different predictor sizes. You can write the program in whatever language you like; although we are going to ask for a printout of your source code, to end result of this assignment are the experimental results (and not the program source code).

Trace and Trace Format

The trace we're giving you is a trace of 16 million conditional branches. These are the conditional branches from an execution of the program GCC (Gnu C Compiler) from the SPECint2000 benchmark suite. As unconditional branches are always taken, they are excluded from the trace. Each line of the trace file has two fields. Below are the first four lines of a trace file:

3086629576 T
3086629604 T
3086629599 N
3086629604 T

The first field is the address of the branch instruction. The second field is the character "T" or "N" for branch taken or not taken.

The trace is available as an 8.6MB gzip compressed file (branch-trace-gcc.trace.gz) or as an 190MB uncompressed text file (branch-trace-gcc.trace). If you wish to verify that the files downloaded correctly, the MD5 sums are 79d52827a9efbb5210a0190b976c20b9 and fd294ab2fc13c2a8e6e7a9b9297c9f29, respectively.

I suggest downloading only the compressed format and processing the compressed file directly. Under Linux this can be done by using zcat to pipe the uncompressed trace into a program reading from standard input. For example: zcat | your-program. Another option is to use zlib (for C or C++) or java.util.zip (for Java) to directly open and process the compressed file. Finally, you can always just use the uncompressed version of the trace.

To help you debug your predictors, we've provided some annotated prediction results for the first 200 branches for various predictors (see below). Comparing your results to these annotated outputs should help ensure your predictor is working properly.

Question 1 - Static Branch Predictor

Before looking at dynamic branch prediction, we are going to look at simple "static" branch prediction policies of "always predict taken" and "always predict not taken". Write a program to read in the trace and calculate the mis-prediction rate (that is, the percentage of conditional branches that were mis-predicted) for these two simple schemes.

  1. Which of these two policies is more accurate (has few mis-predictions)?
  2. Based on common programming idioms, what might explain the above result?

Bimodal Predictor

The simplest dynamic branch direction predictor is an array of 2n two-bit saturating counters. Each counter includes one of four values: strongly taken (T), weakly taken (t), weakly not taken (n), and strongly not taken (N).

Prediction. To make a prediction, the predictor selects a counter from the table using using the lower-order n bits of the instruction's address (its program counter value). The direction prediction is made based on the value of the counter.

Training. After each branch (correctly predicted or not), the hardware increments or decrements the corresponding counter to bias the counter toward the actual branch outcome (the outcome given in the trace file). As these are two bit saturating counters, decrementing a minimum counter or incrementing a maxed out counter should have no impact.

Initialization. Although initialization doesn't effect the results in any significant way, your code should initialize the predictor to "strongly not taken" so that your results match the example traces that we've given you.

Question 2 - Bimodal Accuracy versus Predictor Size

Analyze the impact of predictor size on prediction accuracy. Write a program to simulate the bimodal predictor. Use your program to simulate varying sizes of bimodal predictor. Generate data for bimodal predictors with 22, 23, 24, 25 ... 220 counters. These sizes correspond to predictor index size of 2 bits, 3 bits, 4 bits, 5 bits, ... 20 bits. Generate a line plot of the data using MS Excel or some other graphing program. On the y-axis, plot "percentage of branches mis-predicted" (a metric in which smaller is better). On the x-axis plot the log of the predictor size (basically, the number of index bits). By plotting the predictor size in terms of number of index bits, the x-axis in essence becomes a log scale, which is what we want for this graph.

Answer the following questions base on the data you collected:

  1. Given a large enough predictor, what is the best mis-prediction rate obtainable by the bimodal predictor?
  2. How large must the predictor be to reduce the number of mis-predictions by approximately half as compared to the better of "always taken" and "always not taken"? Give the predictor size both in terms of number of counters as well as bytes.
  3. At what point does the performance of the predictor pretty much max out? That is, how large does the predictor need to be before it basically captures almost all of the benefit of a much larger predictor.

Prediction using Global Branch History (gshare)

A gshare predictor is a more advanced dynamic branch predictor that uses the history of recently executed branches to predict the next branch. Gshare uses a history register to records the taken/not-taken history of the last h branches. It does this by shifting in the most recent conditional branch outcome into the low-order bits of the branch history register. It them hashes the branch history and the PC of the branch when making predictions.

Prediction. Whereas bimodal uses the lowest n bits of the program counter to index into the table, gshare uses the lowest n of the xor of the program counter and the history register. (Note: in C/C++/Java, the "^" operator is the xor operator). Except for this different index into the table, the prediction mechanism is otherwise unchanged from a bimodal predictor. In essence a gshare predictor with zero history bits is basically just a bimodal predictor.

Training. The counter used for prediction (selected by history ^ program counter) is trained just as in a bimodal predictor.

Initialization. Same as bimodal.

Question 3 - Gshare Accuracy versus Predictor Size

Similar to the experiment in question 2, analyze the impact of predictor size on prediction accuracy for a gshare predictor. Extend your program for simulating bimodal predictors to also simulate gshare predictors. Recall that a gshare predictor's configuration is determined by two numbers: the number of history bits and the size of the predictor table.

Use your program to simulate gshare predictor in 8 history bits of varying sizes (the same sizes as in the previous question). Add the gshare data to the graph you created in the previous question. That is, your new graph will have two lines.

Answer the following questions base on the data you collected:

  1. Given a large enough predictor, what is the best mis-prediction rate obtainable by this gshare predictor with 8 history bits?
  2. At what table size is the gshare predictor generally better than the simple bimodal predictor?
  3. Explain why gshare is sometimes more accurate than bimodal.
  4. Explain why bimodal is sometimes more accurate than gshare.

Question 4 - Gshare History Length vs Prediction Accuracy

The previous question used a gshare predictor with 8 bits of history. But how much history should a predictor use? Use your simulator to simulate gshare predictors with varying history lengths (zero bits of history through 19 bits of history). Simulator two different predictor sizes: 210 and 216 predictor entries. Plot one line for each size. The y-axis is as before. The x-axis is now history length.

  1. What is the optimal history length (the history length with the lowest mis-prediction rate) for each of these two predictor sizes?
  2. What do these results say about picking a history length? Generally speaking, what history lengths are the best performing?

Question 5 - Gshare Accuracy versus Predictor Size Revisited

Generate a new series of data in which the gshare history length is the same as the number of index bits. Plot a line of this gshare data on the graph from question #3. This graph should have three lines: the bimodal data, the gshare with a fixed 8-bit history, and the data for gshare where the history length grows with predictor size.

  1. Looking at this new data, what is new the lowest (best) mis-prediction rate for gshare?
  2. When comparing the new gshare data to bimodal, where is the crossover point between then? (That is, what is the predictor size at which gshare becomes better than bimodal?) How does this crossover point compare to the crossover point of gshare with 8-bit history versus bimodal.
  3. In these experiments, we're looking at a trace of just one program (GCC in this case). Of course, processors run many different programs. Each of these programs might have a different crossover point. What properties of a program might most significantly impact where the crossover point between bimodal and gshare occurs?

Tournament Predictor

From the previous data, you can see that neither bimodal or gshare is always best. To remedy this situation, we can use a hybrid predictor that tries to capture the best of both style of predicts. A tournament predictor consists of three tables. The first and second tables are just normal bimodal and gshare predictors. The third table is a "chooser" table that predicts whether the bimodal or gshare predictor will be more accurate.

The basic idea is that some branches do better with a bimodal predictor and some do better with a gshare predictor. So, the chooser is a table of two-bit saturating counters indexed by the low-order bits of the PC that determines which of the other two table's prediction to return. For the choose, the two-bit counter encodes: strongly prefer bimodal (B), weakly prefer bimodal (b), weakly prefer gshare (g), and strong prefer gshare (G).

Prediction. Access the chooser table using the low-order bits of the branch's program counter address. In parallel, the bimodal and gshare tables are access just as normal predictors and they both generate an independent prediction. Based on the result of the lookup in the chooser table, the final prediction is either the prediction from the bimodal predictor (if the choose counter indicates a preference for bimodal) or the prediction from the gshare predictor (otherwise).

Training. Both the gshare and bimodal predictors are trained on every branch using their normal training algorithm. The choose predictor is trained toward which of the two predictors was more accurate on that specific prediction:

As stated above, the bimodal and gshare tables are trained on every branch, totally independent of the chooser table.

Initialization. The chooser table is initialized to strongly prefer bimodal. The gshare and bimodal tables are initialized as normal.

Question 6 - Tournament Predictor Accuracy

Add support to your simulator for the tournament predictor by adding support for the chooser table. If your code was written in a modular way, you should be able to fully re-use your gshare and bimodal code (and thus avoid re-writing or replicating code).

Let's compare the tournament predictor's accuracy versus the data from its two constituent predictors (using the data from the previous question). For now, let's compare a 2n-counter bimodal (or gshare) versus a tournament predictor with three 2n-counter tables. (We'll make the comparison more fair in terms of a "same number of bits" comparison in a moment.) As above for the gshare predictor, the gshare component of the tournament predictor should use a history length equal to the number of index bits (log of the table size). The graph will have three lines total.

  1. How does the tournament predictor's accuracy compare to bimodal and gshare? Is the tournament predictor successful in meeting its goal?
  2. Does the tournament predictor improve the overall peak (best-case) accuracy? If so, why? If not, what are the benefits of the tournament predictor?

Question 7 - Tournament Predictor Accuracy -- Same Size Comparison

In the previous question, the tournament predictor was given the unfair advantage of having three times the storage. In this question, run another set of experimental data in which all the predictors at the same location on the x-axis have the same storage capacity. To do this, compare a 2n-counter predictor to a tournament predictor with the following three table sizes:

  • Chooser table: 2n-2 counters
  • Bimodal table: 2n-2 counters
  • Gshare table: 2n-1 counters

As 2n-2 + 2n-2 + 2n-1 is equal to 2n, this becomes a fair "same size" comparison.

Add a line to the graph from the previous question with this new "fair tournament" data. The graph now has four lines: bimodal, gshare, tournament, tournament-fair.

  1. Compare the old and new tournament predictor data. What was the impact of moving to the smaller (fairer) table sizes?
  2. Once adjusted to be a fair-size comparison, does the tournament predictor succeed in its goal of being the best of bimodal and gshare?
  3. Given a fixed transistor budget for a branch predictor (say, 4KBs), what predictor approach would you use?


The idea of gshare and tournament (or "combining") branch predictors was proposed by Scott McFarling in his seminal paper published in 1993 entitled: Combining Branch Predictors. I encourage you to look at McFarling's paper. After having completed this assignment, I think you'll find the graphs and other results in this paper familiar.

Test Cases

We're provide you with three annotated results (one for each predictor type) from running the first 200 branches from the trace.

What to Turn In

Hints & Comments