Data-oriented Design

Anton Christoffersson

Created: 2023-05-28 Sun 22:11

1. Introduction

Data-oriented design (DOD) is an alternative approach to object-oriented design (OOD) that emphasizes optimizing data layout to improve performance.
While OOD focuses on encapsulating data within objects, DOD separates data and behavior, and organizes data fields in a way that is optimized for how they are accessed and transformed during runtime.
DOD can lead to significant performance improvements in scenarios with large amounts of data and complex operations.

2. Limitations of Object-Oriented Design

OOD emphasizes encapsulating data within objects, which can lead to poor cache locality and frequent cache misses which cause performance issues.

2.1. Limitations of Object-Oriented Design

Simple example of an application performing work on some generic data.

const int ARRAY_SIZE = 100000;

struct FOOD {
  double d1;
  double d2;
  double d3;
};

int main() {
  FOOD dataArray[ARRAY_SIZE];

  for (int i = 0; i < ARRAY_SIZE; i++) {
    dataArray[i].d1 = i;
    dataArray[i].d2 = i;
    dataArray[i].d3 = i;
  }

  for (int i = 0; i < ARRAY_SIZE; i++) {
    dataArray[i].d1 += sqrt(dataArray[i].d1);
    dataArray[i].d2 += sqrt(dataArray[i].d2);
    dataArray[i].d3 += sqrt(dataArray[i].d3);
  }
}

3. Principles of Data-Oriented Design

DOD emphasizes optimizing data layout to improve memory locality leading to better cache efficiency.
This involves separating data and behavior, organizes data fields in a way that is optimized for how they are accessed and transformed during runtime.

3.1. Principles of Data-Oriented Design

Modified example code

const int ARRAY_SIZE = 100000;

struct FDOD {
  double d1[ARRAY_SIZE];
  double d2[ARRAY_SIZE];
  double d3[ARRAY_SIZE];
};

int main() {
  FDOD dataArray;

  for (int i = 0; i < ARRAY_SIZE; i++) {
    dataArray.d1[i] = i;
    dataArray.d2[i] = i;
    dataArray.d3[i] = i;
  }

  for (int i = 0; i < ARRAY_SIZE; i++) {
    dataArray.d1[i] += sqrt(dataArray.d1[i]);
  }
  for (int i = 0; i < ARRAY_SIZE; i++) {
    dataArray.d2[i] += sqrt(dataArray.d2[i]);
  }
  for (int i = 0; i < ARRAY_SIZE; i++) {
    dataArray.d3[i] += sqrt(dataArray.d3[i]);
  }
}

3.2. Principles of Data-Oriented Design

==OOD==
OOD update took avg 1316 microseconds.
D1  miss rate:       7.0% (    6.5%     +     7.7%  )
==DOD==
DOD update took avg 844 microseconds.
D1  miss rate:       4.8% (    5.1%     +     4.1%  )

3.3. Principles of Data-Oriented Design

Spec of the different memory acces speeds of 11th gen Intel of cpus:

Type	Cycles
Register	0.5
L1	2-3
L2	6-8
L3	14-22
DRAM	90-100

4. Benefits of Data-Oriented Design

DOD can improve the performance of an application by optimizing the way data is organized and accessed, reducing cache misses and improving CPU utilization.
DOD can make code simpler and easier to maintain by separating data from behavior and minimizing dependencies between components.
DOD can makes it easier to split the work of processing data across multiple processors or threads by not locking data to objects.