IRJET- Deep Learning Model to Predict Hardware Performance

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 05 Issue: 08 | Aug 2018 www.irjet.net p-ISSN: 2395-0072
Deep Learning Model to Predict Hardware Performance
Shrreenithi Srinivasan1, Avik Satrughana Nayak2
1BE Computer Science, Anna University, Chennai, India
2BE Computer Science, Biju Patnaik University, Bhubaneswar, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract – The goal of the project is to harvest the power
of machine learning algorithms and come up with a novel
approach towards predicting the best possible combination
of hardware features for a computer system, (e.g.
microprocessor/memory specs, Software specs etc.) such
that the baseline and peak scores measured on a certain
benchmark program (e.g. CINT 2006 of SPEC benchmark
program) is maximized. The problem involves predictive
analysis and optimization of hyper parameter for high
accuracy output. The baseline score is the target value. The
hardware features are the independent variables and the
baseline is the dependent variable.
Key Words: Multi Linear Regression, Pandas, Sklearn
package, Backward Elimination, Cross Validation,
Logistic Regression, One hot Encoding, Forward
Selection.
1. INTRODUCTION
An application that is able to predict hardware
performance and analysis with Deep Learning (DL)
algorithms, could drastically improve the performance/
cost curve of a system of machines. An analysis is needed
to verify the result with these expectations. Doing it
manually takes time and this void is filled by a model
which predicts part of the analysis which is the
performance that can be expected with a set of input
configurations. Performing this type of analysis by hand
can require a day or two to complete per benchmark. Thus
a Machine Learning (ML) model is used to do robust
analysis and predict results in a fraction of time. The Deep
Learning model takes input data such as CPU type,
frequency, number of cores, memory size and speed, flash
or disk architecture, network configuration that correlates
against the corresponding performance and system
response. For this project, SPEC CPU 2006 and SPEC CPU
2017 (Reference: 1) will be selected from the
PerfKitBenchmarker suite of industry standard cloud
platform benchmarks. These two Spec benchmarks are the
most popular performance tools in repository which will
be downloaded to PostGresQl for Deep Learning analysis.
It contains SPEC’s next-generation, industry-standardized
CPU intensive suites for measuring and comparing
intensive performance, stressing a system’s processor,
memory subsystem and compiler. Once a Deep Learning
model is chosen and fully implemented, it will have the
ability to infer a score from a given hardware
configuration or infer a hardware configuration from a
given score. This Deep Learning model is used to infer a
benchmark score given a hardware configuration.
2. RELATED WORK
The Deep Learning project is a green field project given by
Flex Cloud Labs to students to explore and implement
machine learning models. Thus there is no previous work
done in the domain as per our investigation.
3. METHODOLOGY
The approach involves training a Deep Learning model
that takes input as the machine parameters and gives the
baseline as output and also maximizes accuracy. To arrive
at an accurate model, we followed a scientific approach
towards optimizing a multi-variable problem involving
following steps.
3.1 Exploratory Data Analysis (EDA)
This is the process of developing an intuition of the input
dataset, analyzing the distribution of mean, median and
standard deviation of scores (in our case the baseline
scores of machine performance), and coming up with an
initial heuristics based model that provides some result
(which need not be accurate), but works within the
framework of existing models.
3.2 Data cleanup
Perhaps the most challenging and major portion of
developing a machine learning model is cleaning up the
input data set. Some studies show that more than 80% of
the work involving the development of Machine Learning
(Reference: 2), involves data cleanup. This is a critical and
often overlooked step, and is of utmost importance since
noise and invalid data will create model parameters that
have little to no correlation to the expected model. We
spent a significant amount of effort in data cleanup using
approaches we explain in the next section.
3.3 Model Fitting
Once we are confident that the dataset has been
sufficiently organized, it is now time to fit a model into our
data set. This requires that we split our dataset into test
© 2018, IRJET | Impact Factor value: 7.211 ISO 9001:2008 Certified Journal | Page 1640

© 2018, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 1641
set and training set and use the training set to generate a
model, and use the test set to measure the accuracy of our
model and fine tune its parameters if necessary. Models
used in predicting hardware performances are one hot
encoding using logistic regression, multi-linear regression
using label encoding, backward elimination, forward
selection and cross-validation.
4. IMPLEMENTATION OF THE DL MODEL
Jupyter notebook, Numpy, Pandas, Sklearn Package and
Python programming language are used to implement this
work. Implementation of deep learning project is done as
follows:
 Data pre-processing (Clean-up, Encoding, Feature
analysis)
 Changing test and training data size
 Algorithm selection & model experimentation (Linear
Regression, Lasso, Logistic etc.)
 Model tuning
 Experiment with Cross validation, Backward
elimination, Forward selection.
 Experimenting with different permutation of above
mentioned factors.
In the area of data analysis, we explored the relationship
between various parameters like hardware, vendor,
system, num_cores, num_chips, processes, num_of_cores
per_chip, auto parallelization, num_of_threads_per_core,
cpu-orderable, processor characteristics, base pointer size,
peak pointer size, first level cache, second level cache,
other cache, memory, operating system, file system on
baseline score. As part of the data analysis we did lot of
research on the datasets and gained insight about the
extent of dependencies with respect to vendor (Figure
4.1), num_cores (Figure 4.2), system (Figure 4.3),
processor (Figure 4.4) and speed (Figure 4.4).
Fig - 4.1: Study on the distribution of different vendors
with respect to baseline score
Fig - 4.2: Study on the distribution of num_cores with
respect to baseline score
Fig - 4.3: Study on the distribution of system with respect
to baseline score
Fig - 4.4: Study on the distribution of processor with
respect to baseline score

Fig - 4.5: Study on the distribution of speed with respect to
baseline score
After implementation, we came up with the clarification
that, cross validation score is getting nearer to training
score with respect to number of datasets (Figure 4.6)
Chart-4.6: Result of cross validation score with respect to
training score
5. RESULTS
We explored different machine learning models such as,
multinomial regression, linear regression, logistic
regression, multi linear regression, multi linear lasso
regression and implemented them to preprocessed dataset
(preprocessing involves data cleaning, encoding dataset
using one hot encoding and label encoding) and also used
various machine learning algorithms like backward
elimination, forward selection (Figure 5.1) and found out
the different results which are mentioned below:
 Using Backward Elimination (all features- model
linear regression) – 80.82%
 Using Forward Selection (all features- model linear
regression) - 75.3%
 Using Cross Validation (only continuous feature- multi
linear regression) - 80.77%
 Using Cross Validation (only continuous feature- multi
linear lasso regression) -78.1%
 Multi Linear Regression (only continuous feature)-
77.64%.
 One Hot Encoding using Logistic Regression
(categorical & all continuous feature) – 50.8%
Fig-5.1: Results of models in predicting score
6. CONCLUSION
Key takeaways in this analysis of deep learning model for
predicting hardware performance are listed as follows:
 Data preprocessing is the main and very important
task when handling with more number of data in the
dataset.
 For Datasets that consists of numeric, alpha numeric
and strings, it is better to use label encoding and one
hot encoding to convert all the alpha numeric and
strings to numeric before passing on to the model. In
this project, both label and one hot encoding is used to
get the better results.
 It is always better to test with different regression
models to check the accuracy and compare the
performance. In this project, we have used linear
regression, logistic regression, multi linear lasso
regression.
 Backward elimination and forward selection helped in
getting good accuracy.
 Best accuracy can be achieved by using backward
elimination and linear regression model with encoding
the features.
ACKNOWLEDGEMENT
We would like to thank our Professor Ms. Vijayalakshmi
for her continuous guidance and encouragement extended
to me. We also like to express our sincere gratitude to the

department for the support and providing the necessary
materials.
REFERENCES
[1] Ould-Ahmed-Vall, E., Doshi, K. A., Yount, C., & Woodlee,
J. (2008). Characterization of SPEC CPU2006 and SPEC
OMP2001
[2] https://www.ibm.com/blogs/bluemix/2017/08/ibm-
data-catalog-data-scientists-productivity/
[3] H. Shakouri |G.R. Nadimi | F.Ghaderi Year 2007: Fuzzy
linear regression models with absolute errors and
optimum uncertainity
[4] Li Zhang | Ruizhen Wu | Yintang Yang Year 2013: A
high-speed and low-power synchronous and
asynchronous packaging circuit based on standard
gates under four-phase one-hot encoding
[5] Fort Lee, N.J (Business Wire): ScaleMP February 22,
2016
https://www.businesswire.com/news/home/201602
22005378/en/ScaleMP-Ranks-No.-1-Standard-
Performance-Evaluation

IRJET- Deep Learning Model to Predict Hardware Performance

More Related Content

What's hot (17)

Similar to IRJET- Deep Learning Model to Predict Hardware Performance (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Deep Learning Model to Predict Hardware Performance