GPU as resources

AI-related resources such as accel­er­a­tion devices need to be iden­ti­fied to pro­vi­sion AI microservice with Fog Plat­form. There are vari­ous types of accel­er­a­tion devices, even an ARM-based SoC is now can be used for edge resources. DECENTER plat­form sup­ports those GPU nodes, either on cloud or edge, to make it able to deliv­er the intel­li­gence with respect to the archi­tec­ture and devices it’s going to run.

QoS-aware orchestration

One of DECENTER’s key innov­a­tions is focused on the deploy­ment phase, par­tic­u­larly when the soft­ware engin­eer has to select an appro­pri­ate IaaS offer to (re)deploy a con­tain­er­ized AI meth­od in the edge-to-cloud con­tinuum. The Qual­ity of Ser­vice (QoS) and Non-Func­tion­al require­ments (NFRs) of AI meth­ods vary greatly, which must be addressed in the deploy­ment phase. For example, for a spe­cif­ic AI meth­od, it may be neces­sary to obtain a high-speed CPU pro­cessor, a large amount of memory, low latency, high band­width, a spe­cif­ic geo­loca­tion for deploy­ment, achieve low oper­a­tion­al cost or sim­il­ar, whilst anoth­er meth­od may require high-speed GPU pro­cessor. There­fore, the prob­lem of select­ing an appro­pri­ate com­pu­ta­tion­al resource, that is, an infra­struc­ture and con­fig­ur­a­tion scheme where a microservice will be deployed and run is not at all unim­port­ant. The more cri­ter­ia used, the high­er the com­plex­ity of the decision-mak­ing pro­cess.

There­fore, with­in the DECENTER pro­ject a new meth­od is being researched and developed that can be used by soft­ware engin­eers to auto­mat­ic­ally rank a list of cloud deploy­ment options for their microservices. The goal of this innov­a­tion con­tends that form­al QoS assur­ances in the applic­a­tion deploy­ment phase can be provided with a new prob­ab­il­ist­ic decision-mak­ing meth­od. Our approach relies on the the­ory and prac­tice of stochast­ic Markov mod­els. Hence, the object­ives of this work have been set out to devel­op:

  • a Markov-based prob­ab­il­ist­ic decision-mak­ing approach, which offers the soft­ware engin­eer a set of optim­al IaaS that would sat­is­fy a spe­cif­ic QoS of the applic­a­tion;
  • an equi­val­ence clas­si­fic­a­tion approach of avail­able cloud deploy­ment options to reduce the com­pu­ta­tion­al time of the decision-mak­ing pro­cess,
  • a design and imple­ment­a­tion of the new meth­od with­in in DECENTER to facil­it­ate (re)deployment oper­a­tions.

Cross-border data-management

In the Big Data era, the vari­ety, velo­city, volume and oth­er aspects of data are prom­in­ent and are cur­rently being addressed by many research­ers world-wide. With­in the DECENTER pro­ject, we focus on the cross-bor­der data-man­age­ment of the DECENTER use cases. In this con­text, a basic Big Data pipeline starts from a cam­era that provides a raw video stream. The AI pro­cessing models/methods are dis­trib­uted to vari­ous resources, which means that we need to deal with dif­fer­ent input and out­put data streams. In the final stage, an out­come of the AI pro­cess takes the format of a struc­tured file with spe­cif­ic inform­a­tion derived from the video-stream, for example, the iden­tity of the per­son shown on the video.

The key innov­a­tion is spe­cif­ic cross-bor­der data man­age­ment mech­an­isms that enable the par­ti­cip­at­ing entit­ies con­trol all aspects of the data trans­port and man­age­ment when it comes to their admin­is­trat­ive domains. Appar­ently, the require­ments extrac­ted by spe­cif­ic parties tak­ing part in a data man­age­ment scen­ario may be too hard in which case it may prove impossible to estab­lish the required qual­ity of cross-bor­der data man­age­ment and trans­port. For instance, a futur­ist­ic European reg­u­la­tion may require cer­ti­fic­a­tion for deal­ing with sens­it­ive private data of all cloud pro­viders that pro­cess per­son­al­ised AI models/methods, anoth­er futur­ist­ic Korean reg­u­la­tion may require to pro­cess sens­it­ive data of Korean cit­izens only on hard­ware resources that are cap­able of using strong secur­ity mech­an­isms (e.g. SGX chips). Fur­ther­more, new GDPR-like European legis­la­tion may require to empower the cit­izen with an abil­ity to approve or dis­ap­prove the use of her/his own per­son­al­ised AI mod­el in spe­cif­ic set­tings, such as at an air­port in Korea. In such case, the cross-bor­der data man­age­ment mech­an­isms must be designed in a way to allow strict assess­ment and applic­a­tion of the user’s pref­er­ences in the spe­cif­ic con­text where the use of AI will be required.