This page collects a set of videos and screencasts related to the either features demonstration or demo of the BarbequeRTRM framework.
Features carousel showing some interesting BarbequeRTRM features.
The demo videos released for the H2020 MANGO project review (http://www.mango-project.eu/).
MANGO boot with test application execution
Early MANGO platform integration demo. The current version includes two “PEAK” many-core processors (based on MIPS). The BarbequeRTRM is responsible for starting the PEAK-side runtime (PEAKos) and enumerates the computing resources. A demo application (matrix multiplication) is then launched with the BarbequeRTRM allocating one of the two processors for the execution.
The BarbequeRTRM managing a deeply heterogeneous HPC prototype system
By exploiting the integration with the MANGO library, the BarbequeRTRM can transparently deploy the application kernels and buffers on the most suitable computing devices and memory nodes. In this video, we see the BarbequeRTRM managing a “MANGO” platform, featuring PEAK multi-core processors (UPV Universitat Politècnica de València ), NUPLUS multi-core SIMD processor (CeRICT) and DCT hardware accelerator (UNIZG University of Zagreb), while running 4 parallel applications (3 instances of a media encoder and an instance of an LDPC).
The demo videos released for the FP7 HARPA project review (http://www.harpa-project.eu/).
CPU thermal control on an ARM big.LITTLE Octa-core CPU
The scenario includes four instances of an image processing application provided by Henesis (http://www.henesis.eu/) managed by the BarbequeRTRM exploiting the TEMPURA policy (version 1). The CPU temperature is kept under a configurable safety threshold.
Performance-aware CPU resource allocation on multi-processor systems
The scenario includes four concurrently running instances of a flood prediction model, monitoring four areas. The application is provided by IT4Innovations (http://www.it4i.cz/). The BarbequeRTRM allocates the CPU resources according to the run-time varying performance requirements of each instance.
Surveillance System Demo
This is an example of a BarbequeRTRM-managed multi-camera surveillance system which is capable of resource reshuffling according to target proximity to an engaging zone. In this scenario, four instances of an OpenCL coded MultiView application are spawned to process four pairs of stereo video streams. The more a moving subject is near to a stereo-camera the more resources are assigned to that stream, thus enabling for an increased frame-rate. The scene has been created using Blender, the Open Source 3D modeling tool.
Runtime Tunable MultiView - University Boot @ DATE13
This demo shows the results of combining application adaptivity capabilities to system-wide resource management. By exploiting a DSE-profiled application (MultiView) and the run-time resource manager BarbequeRTRM, we highlight the benefits of this approach:
Multi-Video Playback Controlled by BarbequeRTRM
In this demo, four OpenCV applications are started concurrently. They are multiple instances of the same application code, competing among them on the usage of the available resources, which are represented by a single CPU since other resources are allocated to high-priority applications. Each video decoder receives a fair amount of 25% of CPU time. Each application runs a QoE run-time manager policy which tries to keep an actual 23 [FPS] rate. After a while, a FAST feature extraction function is enabled on one of the video decoders. This kind of processing, while still matching the frame-rate target, requires more resources, which are asked to the BBQ system-wide resource manager, thus triggering a new resource partitioning among the four (equal priority) video streams. The result is a scheduling decision where the FAST enabled application is given more resources (i.e. processing time, 50%) which are reclaimed from other applications within the same priority level. Moreover, these applications scale down their resolution in order to keep in pace with the frame-rate goal. This demo shows that the BarbequeRTRM is able to:
Face Detection on Android
In this demo, the BarbequeRTRM has been deployed - as a native component - to Android OS (here running on Jelly Bean 4.2.1). As a master thesis project, it has been developed the glue code needed to let Android Apps (Java coded) exploiting BarbequeRTRM features, through Android NDK and JNI features. This demo shows an Android App (Java coded) which relies on a specific Android Service, extending the BarbequeService, which interacts with the native BarbequeRTRM through a specific library. The face recognition algorithm is just emulated, no real coordinates are being computed. The code is freely available on-line: https://bitbucket.org/atroina/bbquesthormfd/overview within the branch named “bosp_android”.
The master thesis is available in PDF for download, as well as the presentation.
Testing BarbequeRTRM deployed under Android OS, through a Java Application which uses JNI interface to link the Java and the native worlds.
This is the really first boot of the BarbequeRTRM on the SThorm platform demo board.
A similar MultiView application but this time a version coded in OpenCL to run on the STHORM platform by STMicroelectronics.
An example of execution of a test application with a resource allocation policy for the SThorm platform, aiming at assigning computing resources within a given power budget.
This demo shows the Stability & Robustness enforcement features of the BarbequeRTRM framework, targeting both applications requirements (AR) and resources availability (RA)
The goal of the demo is to show the BarbequeRTRM capabilities on:
This demo shows the setup of a generic x86 (4×4 cores) NUMA platform, to give the Barbeque RTRM control on a set of processing resources, and different scenarios of workload management.
This demo shows that the BarbequeRTRM is able to:
In this demo, two SVC decoding are started concurrently but with different priorities: the upper video has high-prio (HR_SVC), while the lower one has low-prio (LR_SVC). The upper left screen show that a different cluster resource is assigned to each decoder.
After a while another low-priority workload (BW) is started, which spawns tree EXC competing on resources access with the SVC decoding application.
This demo shows that the BarbequeRTRM is able to: