| <!-- |
| {% comment %} |
| Licensed to the Apache Software Foundation (ASF) under one or more |
| contributor license agreements. See the NOTICE file distributed with |
| this work for additional information regarding copyright ownership. |
| The ASF licenses this file to you under the Apache License, Version 2.0 |
| (the "License"); you may not use this file except in compliance with |
| the License. You may obtain a copy of the License at |
| |
| http://www.apache.org/licenses/LICENSE-2.0 |
| |
| Unless required by applicable law or agreed to in writing, software |
| distributed under the License is distributed on an "AS IS" BASIS, |
| WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| See the License for the specific language governing permissions and |
| limitations under the License. |
| {% endcomment %} |
| --> |
| |
| # Sparse block performance comparison |
| |
| This file contains a short evaluation of different sparse blocks. We included one dense block (DenseBlockFP64) as a baseline. |
| |
| ## Execution time |
| We measured the performance of two matrices being multiplied, both of the corresponding block type. |
| Before each measurement, we performed `30` warmup runs (performing `30` matrix multiplications to random matrices without measuring). |
| Then, we measured the delta nanos of the operation below `100` times and took the average. |
| For this, we selected a block size of `1024x1024`. |
| |
| ```java |
| MatrixBlock m3 = m1.aggregateBinaryOperations(m1, m2, |
| new MatrixBlock(), InstructionUtils.getMatMultOperator(1)); |
| ``` |
| |
|  |
| |
| ## Storage |
| |
| The image below shows the memory consumption of each block type for different sparsities. This was evaluated using blocks with 1024 rows and 1024 columns. |
| |
|  |
| |
| We also evaluated on different matrix shapes by fixing the number of rows and manipulating the number of columns. |
| These tests were repeated for different sparsities. |
| |
|  |
| |
|  |
| |
|  |
| |
|  |
| |
|  |
| |
| To get more meaningful results, the next evaluation fixes the number of entries in a matrix to `1024x1024` and adapts the column-to-row `ratio` (and vice versa) accordingly. |
| We define the ratio as `cl = (1 + ratio) * rl` if `ratio >= 0`, otherwise as `rl = (1 - ratio) * cl`. |
| If we have for example a `ratio=1`, this means that there are twice as many columns as rows. If the `ratio=-1`, this means that there are twice as many rows as columns. |
| Each test with a specific `ratio` was repeated `10` times with different random matrices and we took the average memory consumption. |
| The images below show our results for different sparsities. |
| |
|  |
|  |
|  |
|  |
|  |
|  |
|  |
|  |
|  |
|  |
|  |
|  |
