2018 IEEE High Performance
Extreme Computing Conference
(HPEC ‘18)
Twenty-second Annual HPEC Conference
25 - 27 September 2018
Westin Hotel, Waltham, MA USA
Stripmap SAR Pulse Interleaved Scheduling
Stripmap SAR is a radar mode used to image terrain from an airborne platform; it does so by transmitting and receiving a series
of electromagnetic pulses. Pulse interleaving attempts to execute two or more stripmap tasks simultaneously by transmitting
pulses for additional tasks while waiting for the pulses from other tasks to return, where a task is simply an area on the ground to
image. Other research has done this by dividing the front-end of the radar into separate sections and aiming the energy into
different beams, each pointing in a different direction or devoted to a separate area to image. This thesis focuses on utilizing the
same beam for pulse interleaved scheduling. This research identifies a method for creating a schedule using pulse interleaving
when given a set of stripmap tasks. Scheduling is not done in real time, but is instead done before the schedule will be
executed. Interleaving is performed on multiple tasks overlapping in execution time in the following way. The PRFs of the tasks
are altered within their allowable limits so that they match. Then, the sent pulses and return envelopes, the times when the
pulses might return based on the aircraft's distance to the target and its dimensions, are separated temporally by adding delays,
where needed, to the tasks' transmitted pulses. Doing this removes the possibilities of transmitting a pulse for one task while
receiving the pulse from another task, needing to transmit for multiple tasks at the same time, or receiving pulses from multiple
tasks at the same time. This process allows multiple tasks to be scheduled to execute in the same block of time. To compare
the results of the interleaved scheduler, several greedy algorithms in which no interleaving is permitted were also created. In
every case, the interleaved scheduler significantly outperformed the greedy algorithms. Several situations were created to
simulate different flight conditions as well as radars with varying duty factor limits and output power capability. The performance
of the interleaved scheduler was consistent in all situations. This research demonstrates the benefits of pulse level interleaving
when scheduling multiple stripmap SAR tasks in a short period of time.
Tangram: Colocating HPC Applications with Oversubscription
In a cluster that is shared by many users, jobs often need to wait in the queue for a significant amount of time. Much research
has been done to reduce this time with scheduling, including aggressive back-filling strategies and sharing nodes among
different jobs. Although most resources are shared to some extent in HPC clusters, it is somewhat surprising that a well-known
technique used on commercial clouds, i.e., oversubscribing nodes so that CPU cores are shared among jobs, is rather rare. This
is partially due to concerns about interference. This paper presents Tangram, a framework for colocating applications in HPC
clusters. Tangram uses prior knowledge of applications, such as whether they are I/O or CPU intensive, to predict whether
potential colocations improve overall performance. To predict with sufficient accuracy, Tangram uses a combination of
performance counter measurements, knowledge of past colocation performance, and machine learning. We show that Tangram
can choose colocations to reduce makespan by 19% on average and by 55% in the best case, while limiting the performance
degradation caused by colocation from 1598% to 26% in the worst case.
Experiments on Data Center Participation in Demand Response Programs
Regulation service reserves (RSR) are among the demand response programs in emerging power markets. These programs are
beneficial to the stability of the power grid. RSR requires participants to regulate their power consumption in a near-real-time
manner following a signal broadcast every several seconds, and in return, the participants' electricity cost is reduced as a
reward. Data centers are good candidates to participate in RSR owing to their flexibility in managing power consumption and
computational demand. In this paper, we demonstrate and evaluate data center participation in RSR through experiments on a
13-server cluster. We implement two regulation policies: a policy that primarily tracks the given power signal closely, and another
one that prioritizes quality-of-service (QoS) of the workloads running on the data center. This paper, with policies achieving
acceptable accuracy and good performance even on a small cluster, indicates a promising future where full-scale data centers
participate in demand response programs.
Server-class devices for Space Time Adaptive Processing
Space-Time Adaptive Processing (STAP) is known to efficiently reduce the effects of clutter and jamming. Computing the
adaptive weights in real-time is an intensive process; computational burden can be reduced by selecting computations most
suitable for the processing technology. Even if the most suitable approach is selected for a high-performance multi-channel STAP
system, the processing demand can be substantial. As such, previous generations compute solutions were unable to meet size,
weight and power (SWAP) requirements, making STAP difficult to deploy. A third-order Doppler-factored STAP provides
performance approaching a fully adaptive system. This approach efficiently suppresses clutter and interference. The processing
demand for this algorithm can be high, especially for a high channel count. The number of operations required to adaptively
compute weights increases exponentially which in this case means that a doubling of either channels (L) or processing order (Q)
will quadruple the operation rate. This study will examine a system with 22 channels. This paper examines if recent advances in
technology focusing on Intel Xeon D/E with AVX2 and Intel Xeon SP with AVX512 can make it realistic for embedded applications
to perform high-order STAP while still meeting size, weight and power requirements.
Dynamic Deployment of Communication Applications to Different Hardware Platforms usingOntological
Representations
We consider the problem of mapping communi- cations applications to heterogeneous hardware platforms dy- namically based
on available hardware. Our approach uses an ontology to specify different communications designs and their implementation in
software and hardware. Making use of the ontology and rules, we automatically generate implementations of different
communications applications for FPGA hardware, CPU software, or a hybrid design that mixes the two. Designs are spec- ified in
terms of “tasks” (or processing elements) and “conduits” (or connectors) for data transfer between tasks. Implementations use a
library based approach where tasks are pre-designed for different platforms. Some tasks, such as controllers, are generated at
run time and compiled for the appropriate target hardware. This style of specification-based deployment supports the easy
migration between target hardware platforms. We present a method that uses ontological definitions for automatic code
generation that allows for dynamic deployment of the application to different target hardware. We demonstrate this approach on
an application consisting of two tasks: (1) spectrum sensing and (2) transmission. An automatically generated state machine
implements the control flow, i.e., it decides what frequency to transmit on depending on the available spectrum. Our results show
that this approach generates high quality designs with lower effort from the designer.
Parallel Real-Time Computing
10:20-12:00 in Eden Vale A3
Thursday, September 27, 2018
Stripmap SAR Pulse Interleaved Scheduling
Stripmap SAR is a radar mode used to image terrain from an airborne
platform; it does so by transmitting and receiving a series of
electromagnetic pulses. Pulse interleaving attempts to execute two or
more stripmap tasks simultaneously by transmitting pulses for
additional tasks while waiting for the pulses from other tasks to return,
where a task is simply an area on the ground to image. Other research
has done this by dividing the front-end of the radar into separate
sections and aiming the energy into different beams, each pointing in a
different direction or devoted to a separate area to image. This thesis
focuses on utilizing the same beam for pulse interleaved scheduling.
This research identifies a method for creating a schedule using pulse
interleaving when given a set of stripmap tasks. Scheduling is not done
in real time, but is instead done before the schedule will be executed.
Interleaving is performed on multiple tasks overlapping in execution
time in the following way. The PRFs of the tasks are altered within their
allowable limits so that they match. Then, the sent pulses and return
envelopes, the times when the pulses might return based on the
aircraft's distance to the target and its dimensions, are separated
temporally by adding delays, where needed, to the tasks' transmitted
pulses. Doing this removes the possibilities of transmitting a pulse for
one task while receiving the pulse from another task, needing to
transmit for multiple tasks at the same time, or receiving pulses from
multiple tasks at the same time. This process allows multiple tasks to
be scheduled to execute in the same block of time. To compare the
results of the interleaved scheduler, several greedy algorithms in which
no interleaving is permitted were also created. In every case, the
interleaved scheduler significantly outperformed the greedy algorithms.
Several situations were created to simulate different flight conditions as
well as radars with varying duty factor limits and output power
capability. The performance of the interleaved scheduler was
consistent in all situations. This research demonstrates the benefits of
pulse level interleaving when scheduling multiple stripmap SAR tasks in
a short period of time.
Tangram: Colocating HPC Applications with Oversubscription
In a cluster that is shared by many users, jobs often need to wait in the
queue for a significant amount of time. Much research has been done
to reduce this time with scheduling, including aggressive back-filling
strategies and sharing nodes among different jobs. Although most
resources are shared to some extent in HPC clusters, it is somewhat
surprising that a well-known technique used on commercial clouds, i.e.,
oversubscribing nodes so that CPU cores are shared among jobs, is
rather rare. This is partially due to concerns about interference. This
paper presents Tangram, a framework for colocating applications in
HPC clusters. Tangram uses prior knowledge of applications, such as
whether they are I/O or CPU intensive, to predict whether potential
colocations improve overall performance. To predict with sufficient
accuracy, Tangram uses a combination of performance counter
measurements, knowledge of past colocation performance, and
machine learning. We show that Tangram can choose colocations to
reduce makespan by 19% on average and by 55% in the best case,
while limiting the performance degradation caused by colocation from
1598% to 26% in the worst case.
Experiments on Data Center Participation in Demand Response
Programs
Regulation service reserves (RSR) are among the demand response
programs in emerging power markets. These programs are beneficial to
the stability of the power grid. RSR requires participants to regulate
their power consumption in a near-real-time manner following a signal
broadcast every several seconds, and in return, the participants'
electricity cost is reduced as a reward. Data centers are good
candidates to participate in RSR owing to their flexibility in managing
power consumption and computational demand. In this paper, we
demonstrate and evaluate data center participation in RSR through
experiments on a 13-server cluster. We implement two regulation
policies: a policy that primarily tracks the given power signal closely,
and another one that prioritizes quality-of-service (QoS) of the
workloads running on the data center. This paper, with policies
achieving acceptable accuracy and good performance even on a small
cluster, indicates a promising future where full-scale data centers
participate in demand response programs.
Server-class devices for Space Time Adaptive Processing
Space-Time Adaptive Processing (STAP) is known to efficiently reduce
the effects of clutter and jamming. Computing the adaptive weights in
real-time is an intensive process; computational burden can be reduced
by selecting computations most suitable for the processing technology.
Even if the most suitable approach is selected for a high-performance
multi-channel STAP system, the processing demand can be substantial.
As such, previous generations compute solutions were unable to meet
size, weight and power (SWAP) requirements, making STAP difficult to
deploy. A third-order Doppler-factored STAP provides performance
approaching a fully adaptive system. This approach efficiently
suppresses clutter and interference. The processing demand for this
algorithm can be high, especially for a high channel count. The number
of operations required to adaptively compute weights increases
exponentially which in this case means that a doubling of either
channels (L) or processing order (Q) will quadruple the operation rate.
This study will examine a system with 22 channels. This paper
examines if recent advances in technology focusing on Intel Xeon D/E
with AVX2 and Intel Xeon SP with AVX512 can make it realistic for
embedded applications to perform high-order STAP while still meeting
size, weight and power requirements.
Dynamic Deployment of Communication Applications to Different
Hardware Platforms usingOntological Representations
We consider the problem of mapping communi- cations applications to
heterogeneous hardware platforms dy- namically based on available
hardware. Our approach uses an ontology to specify different
communications designs and their implementation in software and
hardware. Making use of the ontology and rules, we automatically
generate implementations of different communications applications for
FPGA hardware, CPU software, or a hybrid design that mixes the two.
Designs are spec- ified in terms of “tasks” (or processing elements) and
“conduits” (or connectors) for data transfer between tasks.
Implementations use a library based approach where tasks are pre-
designed for different platforms. Some tasks, such as controllers, are
generated at run time and compiled for the appropriate target hardware.
This style of specification-based deployment supports the easy
migration between target hardware platforms. We present a method
that uses ontological definitions for automatic code generation that
allows for dynamic deployment of the application to different target
hardware. We demonstrate this approach on an application consisting
of two tasks: (1) spectrum sensing and (2) transmission. An
automatically generated state machine implements the control flow, i.e.,
it decides what frequency to transmit on depending on the available
spectrum. Our results show that this approach generates high quality
designs with lower effort from the designer.
Parallel Real-Time Computing
10:20-12:00 in Eden Vale A3
Thursday, September 27, 2018
HPEC 2018
25 - 27 September 2018
Westin Hotel, Waltham, MA USA
