364 lines
13 KiB
Markdown
364 lines
13 KiB
Markdown
file:: [PositionPaper+_1673092009371_0.pdf](../assets/PositionPaper+_1673092009371_0.pdf)
|
||
file-path:: ../assets/PositionPaper+_1673092009371_0.pdf
|
||
|
||
- AI needs resources at the edge of the network.
|
||
ls-type:: annotation
|
||
hl-page:: 1
|
||
hl-color:: green
|
||
id:: 63b95bc6-c0fb-4112-8cd0-147470f87d71
|
||
- AI applications developmen
|
||
ls-type:: annotation
|
||
hl-page:: 1
|
||
hl-color:: green
|
||
id:: 63b95bdc-9a45-4dfc-a6e2-20fdb715a460
|
||
- ecure execution
|
||
ls-type:: annotation
|
||
hl-page:: 1
|
||
hl-color:: green
|
||
id:: 63b95be7-fa6d-407d-b5c8-cf17fff66383
|
||
- easy deploymen
|
||
ls-type:: annotation
|
||
hl-page:: 1
|
||
hl-color:: green
|
||
id:: 63b95be9-f2e7-4697-afc4-b36580d418d5
|
||
- runtime management and optimization
|
||
ls-type:: annotation
|
||
hl-page:: 1
|
||
hl-color:: green
|
||
id:: 63b95bec-a6eb-474b-8316-cf3f1211b585
|
||
- AI-SPRINT, a recently funded by the European Commission under the Horizon 2020 framewor
|
||
ls-type:: annotation
|
||
hl-page:: 2
|
||
hl-color:: green
|
||
id:: 63b95bff-660d-48d2-8999-d7a6e7d9adde
|
||
- eamlessly design and partition AI applications among the current plethora of cloud-based solutions and AI-based sensor devices
|
||
ls-type:: annotation
|
||
hl-page:: 2
|
||
hl-color:: green
|
||
id:: 63b95e4f-3544-4b8c-aca2-6cd2699389af
|
||
- design and efficient execution of AI applications exploiting resources in the edge-to-cloud continuum such as flexibility, scalability, interoperability, security and privacy,
|
||
ls-type:: annotation
|
||
hl-page:: 2
|
||
hl-color:: green
|
||
id:: 63b95e95-3364-4411-9f48-e4c6b5df725c
|
||
- Provide design and development tools for the implementation of AI applications
|
||
ls-type:: annotation
|
||
hl-page:: 2
|
||
hl-color:: purple
|
||
id:: 63b95eaa-4593-4cb4-a951-1b80bd3b22c0
|
||
hl-stamp:: 1673102447999
|
||
- hide the communications across components and to transparently implement the parallelization of the compute-intensive part of the application possibly exploiting specialized resources
|
||
ls-type:: annotation
|
||
hl-page:: 2
|
||
hl-color:: green
|
||
id:: 63b95eb6-a419-4fe3-8789-570de920d134
|
||
- Deliver tools for secure execution and privacy preservation
|
||
ls-type:: annotation
|
||
hl-page:: 3
|
||
hl-color:: purple
|
||
id:: 63b95ebb-fd4e-40ef-b2e5-14bd9d480d29
|
||
hl-stamp:: 1673102451024
|
||
- Develop a runtime environment for application execution and monitoring that implements policies orchestrating the applications’ execution across the computing continuum
|
||
ls-type:: annotation
|
||
hl-page:: 3
|
||
hl-color:: purple
|
||
id:: 63b95ec6-2c7a-4a9f-ae3c-5d8f82244acd
|
||
hl-stamp:: 1673102454084
|
||
- Provide advanced solutions for AI architecture enhancement allowing iterative refinement of AI application architecture design and deployment
|
||
ls-type:: annotation
|
||
hl-page:: 3
|
||
hl-color:: purple
|
||
id:: 63b95edb-ec85-4385-a90b-08520ffd3b41
|
||
hl-stamp:: 1673102457572
|
||
- farming 4.0
|
||
ls-type:: annotation
|
||
hl-page:: 3
|
||
hl-color:: green
|
||
id:: 63b95ef5-df9e-476b-89ff-254a8d884788
|
||
- maintenance & inspection
|
||
ls-type:: annotation
|
||
hl-page:: 3
|
||
hl-color:: green
|
||
id:: 63b95ef7-e4a4-4335-9d8b-73dea6393da1
|
||
- personalized healthcar
|
||
ls-type:: annotation
|
||
hl-page:: 3
|
||
hl-color:: green
|
||
id:: 63b95efb-ea49-468a-9fe9-446ad3cb8a6b
|
||
- evelop multi-cloud systems including resources at the full computing continuum stack involving classic components and AI models
|
||
ls-type:: annotation
|
||
hl-page:: 3
|
||
hl-color:: green
|
||
id:: 63b95f0d-4833-4799-8d26-5b088e203700
|
||
- wind farm
|
||
ls-type:: annotation
|
||
hl-page:: 3
|
||
hl-color:: green
|
||
id:: 63b95f21-322b-44e0-90dc-af115981d69b
|
||
- C1 and C2 on the drone
|
||
ls-type:: annotation
|
||
hl-page:: 4
|
||
hl-color:: green
|
||
id:: 63b95fae-7710-49b1-9154-bc9d21b1aff0
|
||
- C3 on the edge server
|
||
ls-type:: annotation
|
||
hl-page:: 4
|
||
hl-color:: green
|
||
id:: 63b95fb1-eb30-4521-a3db-9e64898d44d0
|
||
- C4, C5 and C6 are executed on the cloud.
|
||
ls-type:: annotation
|
||
hl-page:: 4
|
||
hl-color:: green
|
||
id:: 63b95fb4-c63c-4260-87c4-2ef2c82f5cb1
|
||
- ast developmen
|
||
ls-type:: annotation
|
||
hl-page:: 5
|
||
hl-color:: green
|
||
id:: 63b95fed-8cfc-4ea8-a071-65172564b582
|
||
- optimization
|
||
ls-type:: annotation
|
||
hl-page:: 5
|
||
hl-color:: green
|
||
id:: 63b95ff6-f62e-400a-a716-eefd0697a57b
|
||
- and managemen
|
||
ls-type:: annotation
|
||
hl-page:: 5
|
||
hl-color:: green
|
||
id:: 63b95ffb-d224-40fa-9ff6-783e18cba2de
|
||
- AI applications
|
||
ls-type:: annotation
|
||
hl-page:: 5
|
||
hl-color:: green
|
||
id:: 63b95fff-551d-4684-bc75-f494a06195d4
|
||
- unning across a cloud-edge computing continuum
|
||
ls-type:: annotation
|
||
hl-page:: 5
|
||
hl-color:: green
|
||
id:: 63b96006-3be4-4f0d-b7b6-538b42f5bfc0
|
||
- (1) annotating AI applications with non-functional/QoS constraints and guide code parallelizatio
|
||
ls-type:: annotation
|
||
hl-page:: 5
|
||
hl-color:: purple
|
||
id:: 63b96013-bb9a-4acc-b07e-5428e5638428
|
||
hl-stamp:: 1673101193167
|
||
- (2) evaluating a priori performance of AI applications both at training and at inference time
|
||
ls-type:: annotation
|
||
hl-page:: 5
|
||
hl-color:: purple
|
||
id:: 63b96019-9267-4e26-9221-0446fa5d52ac
|
||
hl-stamp:: 1673101195126
|
||
- (3) identifying and automating the search for the most accurate deep neural network starting from labeled datase
|
||
ls-type:: annotation
|
||
hl-page:: 5
|
||
hl-color:: purple
|
||
id:: 63b96022-b2b1-4181-b542-232404cbbae8
|
||
hl-stamp:: 1673101197448
|
||
- (4) automatically exploring multiple candidate deployments for the application components maximizing resource efficiency and minimising the cloud usage cost.
|
||
ls-type:: annotation
|
||
hl-page:: 5
|
||
hl-color:: purple
|
||
id:: 63b9602c-4343-48d1-adea-8e2afcc0c6ae
|
||
hl-stamp:: 1673101200142
|
||
- i) intra- and intercomponents parallelism
|
||
ls-type:: annotation
|
||
hl-page:: 5
|
||
hl-color:: green
|
||
id:: 63b96042-9a8b-453b-bdc9-57eafd6844d5
|
||
- ii) constraints on application performance, security, privacy, and underlying resources
|
||
ls-type:: annotation
|
||
hl-page:: 5
|
||
hl-color:: green
|
||
id:: 63b96049-22a2-48f5-8ad0-4c446bcd8ea7
|
||
- iii) deployment alternatives that will be automatically explored by the envisioned solutions.
|
||
ls-type:: annotation
|
||
hl-page:: 5
|
||
hl-color:: green
|
||
id:: 63b9604f-8305-4eca-8161-89d59f86c37d
|
||
- predict the execution time of inference or training tasks of AI models deployed across the computing continuum.
|
||
ls-type:: annotation
|
||
hl-page:: 7
|
||
hl-color:: green
|
||
id:: 63b960bf-8ee8-4f27-994c-346f3556d6cc
|
||
- AI-SPRINT will extend the use of such models to consider AI-based sensors and deep networks partitioned and deployed across computing continua
|
||
ls-type:: annotation
|
||
hl-page:: 7
|
||
hl-color:: green
|
||
id:: 63b960d1-9c7f-4399-ab05-906132063ee3
|
||
- limited ML expertise to train high-quality models specific to their needs also in terms of Quality of Service (QoS) requirements
|
||
ls-type:: annotation
|
||
hl-page:: 7
|
||
hl-color:: green
|
||
id:: 63b960e0-d535-4f55-a848-6e3952d2ba29
|
||
- AI-SPRINT Running Example
|
||
ls-type:: annotation
|
||
hl-page:: 3
|
||
hl-color:: green
|
||
id:: 63b9620c-5222-40f5-afb5-242b358cf5bd
|
||
- internal parallelism
|
||
ls-type:: annotation
|
||
hl-page:: 5
|
||
hl-color:: yellow
|
||
id:: 63b97f7f-9f49-43d4-8145-0b61fb71e001
|
||
hl-stamp:: 1673101185237
|
||
- programming layer
|
||
ls-type:: annotation
|
||
hl-page:: 5
|
||
hl-color:: purple
|
||
id:: 63b98006-1e1a-487c-b9a7-f94688ec2e71
|
||
- https://www.bsc.es/research-and-development/software-and-apps/software-list/compsuperscalar
|
||
ls-type:: annotation
|
||
hl-page:: 6
|
||
hl-color:: purple
|
||
id:: 63b98016-d678-480b-886e-d661cd1b5815
|
||
- PyCOMPSs
|
||
ls-type:: annotation
|
||
hl-page:: 6
|
||
hl-color:: purple
|
||
id:: 63b98073-ff31-4834-a588-d9208de2aceb
|
||
- dentifying the functions to be executed as asynchronous parallel task
|
||
ls-type:: annotation
|
||
hl-page:: 6
|
||
hl-color:: yellow
|
||
id:: 63b98089-4e33-41e8-b227-3d7e6ab3c85e
|
||
hl-stamp:: 1673101471686
|
||
- annotating them with standard Python decorators
|
||
ls-type:: annotation
|
||
hl-page:: 6
|
||
hl-color:: yellow
|
||
id:: 63b9808d-dd15-4e61-95ca-bb104f429333
|
||
hl-stamp:: 1673101473254
|
||
- New annotations will be introduced to allow predicating on components performance and to specify constraints on the target deployment
|
||
ls-type:: annotation
|
||
hl-page:: 6
|
||
hl-color:: purple
|
||
id:: 63b980a9-4c4b-4be8-95a4-62e73a1d94b6
|
||
hl-stamp:: 1673101482951
|
||
- as a parallel framework, PyCOMPSs exploits inter-node and intra-node parallelism and executes neural networks tasks (like Tensorflow or Keras) as external processes.
|
||
ls-type:: annotation
|
||
hl-page:: 6
|
||
hl-color:: purple
|
||
id:: 63b98114-59de-4db0-9f78-c627bc04d79f
|
||
- the design still requires specifying the parameters of a typically large network architecture with several layers and units, and then solve a difficult non-convex optimization problem
|
||
ls-type:: annotation
|
||
hl-page:: 7
|
||
hl-color:: green
|
||
id:: 63b98132-e0ac-4383-aaa1-be5569cbb83c
|
||
hl-stamp:: 1673101620848
|
||
- AI-SPRINT will provide tools for developing learning as a service solution, that, starting from a training set with labelled training examples (images or temporal data series which are of interest for use cases) will automatically identify the most accurate deep neural network which provides execution time guarantees.
|
||
ls-type:: annotation
|
||
hl-page:: 7
|
||
hl-color:: purple
|
||
id:: 63b98142-f9e9-4129-aa93-865aba5900b3
|
||
hl-stamp:: 1673101636290
|
||
- AI-SPRINT will explore solutions where the edge servers or the AI-enabled sensors run the initial layers of the neural network, and then send the output to the cloud to feed the remaining layers and produce the final result.
|
||
ls-type:: annotation
|
||
hl-page:: 7
|
||
hl-color:: purple
|
||
id:: 63b98164-8180-4d6b-8cb6-8934465109a9
|
||
- AI-SPRINT runtime environment will include tools to: i) support the continuous deployment of AI applications;
|
||
ls-type:: annotation
|
||
hl-page:: 8
|
||
hl-color:: purple
|
||
id:: 63b9818e-5338-488b-af75-a529b47b90d1
|
||
- i) support application components concurrent execution reacting to node failures and identifying their optimal placement and resource capacity
|
||
ls-type:: annotation
|
||
hl-page:: 8
|
||
hl-color:: purple
|
||
id:: 63b98199-ac2e-4404-834d-66f3c52e918e
|
||
- iii) trigger automated model retraining, leveraging on solutions for training and retraining at the edge;
|
||
ls-type:: annotation
|
||
hl-page:: 8
|
||
hl-color:: purple
|
||
id:: 63b9819f-57aa-44d5-af6e-72733e67c64e
|
||
- v) optimize the scheduling and assignment of accelerator devices among competing training jobs.
|
||
ls-type:: annotation
|
||
hl-page:: 8
|
||
hl-color:: purple
|
||
id:: 63b981a5-20cd-4c71-879a-c2f85d14dc1e
|
||
- Continuous deployment is the process where development, testing, delivery and progressive deployment of an application is considered as a whole, reducing operation and development costs
|
||
ls-type:: annotation
|
||
hl-page:: 8
|
||
hl-color:: green
|
||
id:: 63b98246-6754-4a3a-b835-6289f9b76bee
|
||
hl-stamp:: 1673101896873
|
||
- This is especially important in a scenario where data is continuously retrieved, and AI models are retrained.
|
||
ls-type:: annotation
|
||
hl-page:: 8
|
||
hl-color:: purple
|
||
id:: 63b98269-1951-422b-adaf-3371d4d59ca7
|
||
hl-stamp:: 1673101932159
|
||
- heterogeneity of the edge laye
|
||
ls-type:: annotation
|
||
hl-page:: 8
|
||
hl-color:: purple
|
||
id:: 63b98271-e42a-4fe7-8e2f-d98c20ff0d30
|
||
- streamlined mechanisms to guide the process of configuring computing resources across the computing continuum
|
||
ls-type:: annotation
|
||
hl-page:: 8
|
||
hl-color:: purple
|
||
id:: 63b9827a-24fa-4c96-bfe1-63281e57a0a6
|
||
- https://www.egi.eu/services/cloud-compute
|
||
ls-type:: annotation
|
||
hl-page:: 8
|
||
hl-color:: purple
|
||
id:: 63b98296-0d83-4c54-b756-97c96e77f101
|
||
- ocker containers to automate deployment and provide the level of isolation needed to enforce performance constraints with minimal overhead.
|
||
ls-type:: annotation
|
||
hl-page:: 9
|
||
hl-color:: purple
|
||
id:: 63b982dc-4503-4bb9-8cf0-85e82531af62
|
||
- OASIS TOSCA templates
|
||
ls-type:: annotation
|
||
hl-page:: 9
|
||
hl-color:: purple
|
||
id:: 63b982e3-6c41-41d2-b7cc-73cc945054d9
|
||
- AI model architecture search
|
||
ls-type:: annotation
|
||
hl-page:: 13
|
||
hl-color:: purple
|
||
id:: 63b98326-ca0e-4376-a724-af4aa045cfbd
|
||
- AI cloud-edge application design
|
||
ls-type:: annotation
|
||
hl-page:: 13
|
||
hl-color:: purple
|
||
id:: 63b9832b-fbe6-468b-9ef7-a330b97080e5
|
||
- CNNs for image recognition
|
||
ls-type:: annotation
|
||
hl-page:: 13
|
||
hl-color:: purple
|
||
id:: 63b9833b-7517-4a2f-8895-0a5f19e40e67
|
||
- multi-objective perspective trading-off accuracy with energy consumption
|
||
ls-type:: annotation
|
||
hl-page:: 13
|
||
hl-color:: purple
|
||
id:: 63b98346-44d5-4240-8fa1-e2d8de446bce
|
||
- for edge device
|
||
ls-type:: annotation
|
||
hl-page:: 13
|
||
hl-color:: purple
|
||
id:: 63b98348-8d81-475f-959b-a73bfac64f4c
|
||
- federated setting
|
||
ls-type:: annotation
|
||
hl-page:: 13
|
||
hl-color:: purple
|
||
id:: 63b9834a-49fd-479a-8090-cfc0a768df53
|
||
- cloud orchestration runtime that deploys complex and customized virtual infrastructures on multiple back-ends
|
||
ls-type:: annotation
|
||
hl-page:: 13
|
||
hl-color:: purple
|
||
id:: 63b98375-330a-4caa-bd89-cc7bcc89a831
|
||
- edge limited capacity
|
||
ls-type:: annotation
|
||
hl-page:: 13
|
||
hl-color:: purple
|
||
id:: 63b98382-9165-4d37-88ad-1f7a9aa820e7
|
||
- several approaches have been proposed to tackle this issue
|
||
ls-type:: annotation
|
||
hl-page:: 13
|
||
hl-color:: purple
|
||
id:: 63b98386-0238-43bc-ab62-32188dd57970
|
||
- AI-SPRINT Objectives
|
||
ls-type:: annotation
|
||
hl-page:: 2
|
||
hl-color:: purple
|
||
id:: 63b98487-4ce8-4ff0-8697-d3b08a66d8e2 |