2. http://www.harness-project.euhttp://www.harness-project.eu/
Introduction
Goal of HARNESS project: allow cloud providers and tenants
take advantage of heterogeneous resources
Key idea: heterogeneous resources as first-class entities
FPGAs, GPGPUs, programmable routers, heterogeneous
storage volumes
Benefits:
offer a wider context to make price/performance trade offs
provide larger scope for resource allocation/optimization
increase application performance
reduce energy consumption and cost profiles
4. http://www.harness-project.euhttp://www.harness-project.eu/
The RTM HARNESS Use Case
1. Generates an image of the sub-surface from seismic
survey data
Processing ~10,000 “shots”
1 shot = 1 acoustic stimulus +
~16 s recording on ~10,000 hydrophones
2. Embarrassingly parallel between shots
3. Can project full RTM runtime from single shot
7. http://www.harness-project.euhttp://www.harness-project.eu/
DEMO: Cloud Tenant View
Profiling - Performance Model Building
1. User uploads application + manifest via web
frontend
2. Platform creates Application Manager which starts
profiling to build performance model
Manifest defines the configuration space:
1 and 3 CPU cores, 2 to 4 DFEs
Each execution runs via CRS + IRMs
3. Tenant can download generated performance model
8. http://www.harness-project.euhttp://www.harness-project.eu/
Step-by-Step Production View
Step1:
User submits
- Application zip
- Application Manifest
- Application performance model
Step2:
Director assesses user’s request,
if ok creates an
Application Manager instance
Step3:
AM requests necessary resources
to CRS
Step4:
CRS requests the resources
through the appropriate IRMs
Step5:
IRM creates instance(s) in the
relative physical resource and
provides appropriate handles to AM
Step6:
AM deploys the application in the
instances created
Step7:
AM, once application up and
running, provides end point to user
9. http://www.harness-project.euhttp://www.harness-project.eu/
Testbed @ Imperial
Axel07
Maia01
Demo Preview
CRS
OpenStack
I
R
M
Instance: RTM
VPCU = 2
DFE = 1
Memory = 1GB
Host = Maia01
Instance type:
Application
Manager
VPCU = 1
Memory = 0.5GB
Host = Axel07
Front
End
Director
• 1 VM: 2VCPUs, 1G RAM,
100MB Disk, 1DFE …
• 1 VM: 1VCPUs, 0.5G
RAM, 200MB Disk, …
SLO
I want run my
zip app with X
performance
and Y budget
10. http://www.harness-project.euhttp://www.harness-project.eu/
Demo: Cloud Tenant View
1. User uploads application + manifest via web
frontend
2. User uploads application performance model
27 configurations: 1 to 4 DFEs, 1GB and 2GB RAM,
10MB/s and 180MB/s storage performance, 1 to 3 CPU
cores
3. User specifies objectives for run
4. Platform indicates whether objectives are satisfiable
5. User triggers execution
6. Application Manager selects resource configuration
and executes process
14. http://www.harness-project.euhttp://www.harness-project.eu/
Conclusion
Goal of HARNESS project: allow cloud providers and tenants
take advantage of heterogeneous resources
Key idea: heterogeneous resources as first-class entities
FPGAs, GPGPUs, programmable routers, heterogeneous
storage volumes
Benefits:
offer a wider context to make price/performance trade offs
provide larger scope for resource allocation/optimization
increase application performance
reduce energy consumption and cost profiles
Editor's Notes
HDD is network connected and slow performance, 10MB/s
About the RTM application:
It’s a storage version, stores snapshots as it’s computing the shots
Huge amount of data to process
The generation of the image from the data is a heavy and costly process
We show here the video of the application
************************
VIDEO: RTMDemoVideo1_HD
************************
Production was shown in the opening demo.
In the profiling phase the systems know nothing about the application performance, so it explores various configurations in terms of performance times and cost
we show here the video of the profiling
************************
VIDEO: y2_rev_demo_profiling.mp4
************************
Considerations with this approach:
As there are various user processes (profile and production), only one can be shown here. Production is chosen
There are many things underneath happening, and simplifying the message might lead to misleading information
E.g. the CRS does many things to abstract the resources to the AM, but this is difficult to convey in the simplified step-by-step view. Saying that CRS allocates the resources may pass the msg that the CRS does a very easy task, which is not
Just mention that some of the components do more than what’s shown here
we show the video of the production phase
************************
VIDEO: y2_rev_demo_production_v1.mp4
************************
We show the video of the resource view first and then the messaging one
************************
VIDEO: crs and msg
************************
This slides highlights where the application is wrt the architecture