Presentation given by Tobias Becker (Maxeler) at the LEGaTO Final Event: Low-Energy Heterogeneous Computing Workshop on 4 September 2020 This event was collocated with FPL 2020

1. The LEGaTO project has received funding from the European Union's Horizon 2020 research and innovation programme under
the grant agreement No 780681
9 September 2020
Infection Research with
Maxeler Dataflow
Computing
LEGaTO Final Event: Low-Energy Heterogeneous
Computing Workshop
Tobias Becker
Maxeler Technologies

2. Low-Energy Heterogeneous Computing Workshop
Introduction – Biomarkers Analysis
• Biomarkers can serve many unique purposes including :
− screening for early signs of disease, or monitoring of progression
− monitoring effects of the treatments
− study organ function
• Various data types
− measurable molecules found in body tissues, cells and fluids
− MRI
− function tests
− heart rate, blood pressure
• Modern high-throughout technologies can produce lots of data:
micro array, next-gen sequencing, mass spectroscopy
• Goal: Using statistical methods to research effectiveness of drugs,
vaccination strategies and harmfulness of pathogens
2 09/09/20

3. Low-Energy Heterogeneous Computing Workshop
Computational aspects
• High-throughput data from biological experiments are gathered into vectors
whose dimension correspond to the number of measurements
• For statistical analysis, the number of samples (n) must be larger than the
number of biomarker candidates (p) for good classification performance to get a
reliable diagnosis.
• Obtaining samples is expensive
• Basis: Pilot studies
− Small sample size but large number of candidates (short fat data)
− Distinguishing real correlations from random correlations
• Strategy:
− Select top features from thousands that can predict the cases
− Estimating the appropriate sample size to get significant results
3 09/09/20

4. Low-Energy Heterogeneous Computing Workshop
Algorithm: Biomarker Candidate Evaluation
4 09/09/20
Pilot study with small sample size
Global evaluation of single biomarkers
More ``good´´
single
biomarkers
than expceted?
Discard/Redesign study
no
Compute
potential of
biomarker
combinations
& estimate
required
sample size to
validate
combinations
yes
Real Data
Compute goodness
of BCs
(Entropy)
Random Data
Evaluate pilot study
(HiPerMAb)
Estimate p-values
Maxeler DFE
Implemented in R
Time consuming: runs hours on typical datasets

5. Low-Energy Heterogeneous Computing Workshop
Optimisation
• Goals:
− Decrease simulation runtime
− Enable to handle larger data sets and more
complex problems
− Decrease energy consumption
• Target:
− Maxeler dataflow computer with MAX5 DFE
− Also works on Xilinx Alveo and Amazon EC2 F1
• Optimisation plan:
− Manual: Port R code to C and MaxJ
− Optimise using LEGaTO toolflow: Port to OmpSs
and MaxJ
5

6. Low-Energy Heterogeneous Computing Workshop
PCI
Express
Manager
DFE
Memory
MyManager (.maxj)
x
x
+
30
x
Manager m = new Manager();
Kernel k =
new MyKernel();
m.setKernel(k);
m.setIO(
link(“x", CPU),
m.build();
link(“y", CPU));
MyKernel(DATA_SIZE,
x, DATA_SIZE*4,
y, DATA_SIZE*4);
for (int i =0; i < DATA_SIZE; i++)
y[i]= x[i] * x[i] + 30;
Main
Memory
CPU
CPU
Code
Host Code (.c)
MaxCompiler: Application Development in MaxJ
6
SLiC
MaxelerOS
DFEVar x = io.input("x", dfeInt(32));
DFEVar result = x * x + 30;
io.output("y", result, dfeInt(32));
MyKernel (.maxj)
int*x, *y;
x
x
+
30
y

7. Low-Energy Heterogeneous Computing Workshop
Practical Acceleration Process in MaxJ
Focus profiling on loop structure, not function calls
init_run : 66.92s CPU 66.92s WALL ( 1 calls)
electrons : 3116.79s CPU 3116.79s WALL ( 1 calls)
Called by electrons:
v_of_rho : 17.91s CPU 17.91s WALL ( 51 calls)
c_bands : 2305.58s CPU 2305.58s WALL ( 50 calls)
sum_band : 597.16s CPU 597.16s WALL ( 50 calls)
mix_rho : 2.80s CPU 2.80s WALL ( 50 calls)
Called by c_bands:
h_psi : 962.63s CPU 962.63s WALL ( 684 calls)
g_psi : 25.24s CPU 25.24s WALL ( 582 calls)
cdiaghg : 666.02s CPU 666.02s WALL ( 682 calls)
Called by h_psi:
add_vuspsi : 79.00s CPU 79.00s WALL ( 684 calls)
General routines
calbec : 56.10s CPU 56.10s WALL ( 684 calls)
fft : 3.30s CPU 3.30s WALL ( 560 calls)
fftw : 1156.55s CPU 1156.50s WALL ( 97578 calls)

8. Low-Energy Heterogeneous Computing Workshop
Task-based kernel identification for DFE mapping
• OmpSs identifies “static” task graphs while running
• Annotation of IO and compute help to create DFE task model
• Instantiate static, customized, ultra-deep (>1,000 stages) computing pipelines

9. Low-Energy Heterogeneous Computing Workshop
LEGaTO tools for DFE mapping
9 September 2020
task graph
loopflow graph
• automatically generate loopflow graph
• annotate with key information

10. Low-Energy Heterogeneous Computing Workshop
Results
10
• Application accelerated with Maxeler DFE
− Accelerated Cut Index function (entropy calculation) in Biomarker Evaluation
− Added hardware random number generator to Cut Index accelerator
− Evaluated 10^6 problem size on Jülich testbed
− Validated new toolflow and results
Original R code C++ Cut Index DFE Cut Index
Time[s] 4514 74.8 5.49
Speedup 1(baseline) 60 822
Original R code C++ Cut Index DFE Cut Index
Energy [Wh] 575 9.7 1.3
efficiency gain 1(baseline) 59 443

11. Thanks!