1. Roman Atachiants
Supervised by Dr. Gavin Doherty
Collaborators Dr. David Gregg, Dr. Bérenger Arnaud
Project Title MANYCORE: Understanding Software
Performance on Many-Core Systems
Colour
Photo
6 x 4 cm
Visualising Data Locality Performance
In order to take advantage of the multi-core and many-core
hardware of today and tomorrow, programmers are faced with a
need to parallelize the code to distribute work across multiple CPUs.
User Evaluation
To evaluate the tool,
we have conducted an
experiment with a total
of 33 participants from
industry and academia.
To analyze the results
of the experiment, we
adopted a hybrid quali-
tative and quantitative
approach.
Fieldwork, Taxonomy, Modelling and Validation
While the visualisation itself is being the final step of this research, it
is based on the significant amount of analytical work and an
observational model we’ve created and validated to bridge the gap
The process of paralle-
lization is very complex
and in order to assist
programmers in identi-
fying the performance
of parallel programs re-
lated to poor data loca-
lity, we have designed
an interactive visualis-
ation tool.
3-Step Visualisation
Summary View
Timeline View
Threads View
Our Publications
R. Atachiants, D. Gregg and
G. Doherty. Design
Considerations for
Parallel Performance
Tools. ACM SIG-CHI’14
R. Atachiants, D. Gregg and
G. Doherty. 2015. An
Observational Model for
Identifying Parallel
Performance Problems.
Journal paper submitted and
under revision
R. Atachiants, D. Gregg and
G. Doherty. Visualising
Data Locality Performance
for Parallel Programming.
Submitted for revision.
R. Atachiants. Ph.D.
Thesis: Supporting Visual
Diagnosis of Performance
Problems in Multi-Threaded
Software. [Draft]
Some Results
The participants' correctness in data locality problem identifica-
tion has significantly increased when they used the visualisation.
We received a significant amount of feedback which suggests
that the visualization effectively supports programmers and
reduced the cognitive load of performance problem diagnosis.
Programmers with less than 10
years of experience in the field
rated their diagnosis answers
with significantly more confi-
dence when they used our
visualisation.
between the events and
counters we can collect
and the actual parallel
performance problems.
We have also created a
taxonomy of problems
comprised of 23 parallel
problems and ran two
experiments with the
help of 81 programmers
to validate the taxono-
my and our model.