Final Project - Load Value Prediction

Deadline: March 17 @ 11:59:59PM. This is a hard deadline. Note: Slip days cannot be used for the final project.

Project logistics

Group project / Submission

The final project can be conducted in groups of up to 3 students. You can choose to have a smaller group of 2 or as an individual.
Once you have formed your group join the following Github Classroom assignment and create a repo for your group:
The repository that will be created is empty since this is an open-ended project. You will use this repository to submit your final project with the deliverables below.


  • Report that describes:
    • Your approach towards converging to a load value predictor design (Why did you design it that way?)
    • The design and implementation of your predictor(s) (How did you design it?)
    • Your experimental setup/methodology (How did you test your design?)
    • Evaluation and analysis of your predictor (Does your design work?)
  • Documentation / instructions on how to compile and run your predictor.
    • Include the exact commands that was used to generate your results and to evaluate your predictor
  • Source code
  • Sample traces which you used to evaluate and test See the generating memory load traces section below.

Load value prediction overview

The purpose of load value prediction is to be able to predict the value of a memory load instruction.

Recall that memory accesses can be hundreds or thousands of cycles and can lead to significant pipeline stalls, especially on cache misses.
One unique solution around this is to be able to predict the value of the load instruction and speculatively execute based on the predicted load value.
Later when the actual memory load value returns, we can compare and check to see if our load value was correctly predicted or not.
Similar to our branch misprediction scenario in speculative tomasulo, if our load value is mispredicted we simply rollback by flushing the ROB.
If we correctly predicted, then great! We continue running on and benefited from executing instructions speculatively.

A major reason why value prediction works is that programs tend to exhibit value similarity behaviors where values tend to be similar either spatially (neighboring memory addresses) or temporally (repeated access to the same memory location does not change, for example, a constant value in memory).
If you're curious, you can refer to the following seminal paper by Lipasti, et. el for details: Value Locality and Load Value Prediction

You can abstractly think of value prediction as the figure below:

Given an instruction PC (which represents "What memory load instruction is this?"), a predictor function is able to predict the value that will be read from memory.
There are various examples of simple predictors, such as using the last value that was loaded from this address, the last value + a stride, using global load value information, or using control-flow information.
These predictor functions have the same limitations as our branch predictor, in that it is resource limited and typically based on tables of information.

In addition to the predicted load value, load value predictors also include confidence values, that is, "Should I use the predicted value of this load instruction or not?"
The load value predictor typically learns if a certain load instruction has load value properties that can be extracted or not.
For example, if you have random numbers populated in memory, then a load value predictor should not be used since you'll probably mispredict a lot and end up rolling back a lot which will make your processor slower overall.

Final project requirements

Your goal for the final project is to be able to develop a load value predictor.

Load value predictor requirements

The predictor will use a memory access trace that contains the load PC, the load address and the load value as detailed in the next section. The inputs to the predictor will be:

  • The PC address of the load instruction
  • The memory address to be access by the load

The output of the predictor will be:

  • The predicted value
  • Whether you should use the predicted value or not (the confidence). For simplicity, this output should be binary. (Yes/No)

Load value predictor implementation

There are two options for this final project.

Option 1: Practical load value predictor -- Similar to our class assigment, you can create a table-based predictor that will try to achieve as high of an accuracy as possible. This is completely open-ended so you can choose to try different designs to improve accuracy. For comparison, you are required to implement a simple last value predictor where you simply predict the value that was last loaded for that load PC and use the prediction for every load.

Option 2: Magical machine learning load value predictor -- Go crazy and have fun. Imagine your load value predictor can have infinite resources and compute power and can run a whole ML framework in it. Use Tensorflow, Keras, pyTorch, insert ML framework here, to build your predictor. This is open-ended so you can desgin any type of ML that you find suitable. Note that most likely it will be some kind of reinforcement learning since you have to learn as the program is executing (as you're going through the trace). The goal here is not to perform supervised learning (off-line training using a subset of the trace as a training set and another subset as testing set) since load value predictors in real hardware cannot be pre-trained and must learn on-the-fly.

Generating memory load traces with Pin Dynamic Binary Instrumentation tool

For reference, Pin Manual:

Download and compile

Run the pintool to collect a memory trace

Below is an example command to run the pin tool on the ls command.

../../../pin -t obj-intel64/ -- /bin/ls

Memory trace is output to file pinatrace.out and contains the PC address, read/write to memory, and the memory address.

$ head pinatrace.out 
0x7f2498c8f093: W 0x7ffc630ad988
0x7f2498c8fe90: W 0x7ffc630ad980
0x7f2498c8fe94: W 0x7ffc630ad978
0x7f2498c8fe96: W 0x7ffc630ad970
0x7f2498c8fe98: W 0x7ffc630ad968
0x7f2498c8fe9a: W 0x7ffc630ad960
0x7f2498c8fe9f: W 0x7ffc630ad958
0x7f2498c8feaf: R 0x7f2498cb5e78
0x7f2498c8feb6: W 0x7f2498cb56e0
0x7f2498c8fec7: R 0x7f2498cb6000

Modify pinatrace.cpp to print out load data values only

Since we care only about load value prediction, we can comment out lines 58-65 of pinatrace.cpp which instruments store instructions.

Modify the RecordMemRead() function to printout the value loaded from memory as shown below:

// Print a memory read record
VOID RecordMemRead(VOID * ip, VOID * addr)
    ADDRINT value;
    PIN_SafeCopy(&value, addr, sizeof(ADDRINT));    
    fprintf(trace,"%p: R %p, 0x%016lx\n", ip, addr,value);

Now you can collect load value traces. (Make sure you recompile.)

Example: Make a trace of your assignment 1 program

Run your pintool on your assignment 1 pipesim similar to below. You will likely have to replace the path that is suitable for your environment.

pin-3.18-98332-gaebd7b1e6-gcc-linux/source/tools/ManualExamples$ ../../../pin -t obj-intel64/ -- ~/CS203/assignment1-pipeline-wongdani/pipesim -i ~/CS203/assignment1-pipeline-wongdani/traces/instruction1.txt 

The memory load value traces will be in pinatrace.out. For my pipesim, the output file is 29MB, containing 604K of memory loads with 209K unique memory loads (note there are many duplicate memory load values!)

pin-3.18-98332-gaebd7b1e6-gcc-linux/source/tools/ManualExamples$ ls -lh pinatrace.out 
-rw-r--r-- 1 dwong dwong 29M Feb 21 14:10 pinatrace.out
pin-3.18-98332-gaebd7b1e6-gcc-linux/source/tools/ManualExamples$ wc -l pinatrace.out 
604177 pinatrace.out
pin-3.18-98332-gaebd7b1e6-gcc-linux/source/tools/ManualExamples$ sort pinatrace.out | uniq | wc -l

For your final project, you will need to generate different load value traces to train/test your load value predictor