In our previous blog, we presented a blockchain-reputation system for tackling reliability problems in edge offloading, which captures the long-term performance of edge servers and the protection of reputation information against malicious tampering. There are a few challenges of blockchain-based reputation systems that remain unsolved like a challenge (C1) that does not provide a formal guarantee about the feasibility of offloading decisions which is important for high-reliability applications (C2) difficulty operating in distributed edge environments where reliability levels are highly diverse, (C3) not addressing reliability in terms of edge failures, and (C4) neglecting the impact of long-latency blockchain consensus on latency-sensitive offloading decisions. We present an edge offloading framework called FRESCO, which optimizes performance and reliability in distributed unreliable edge offloading scenarios for latency-sensitive applications (the problem is presented in detail in a blog).

FRESCO Edge Offloading Framework

FRESCO consists of an offloading decision engine and a reputation state manager. An offloading decisions engine is a software unit deployed on a mobile device that decides on which server task is going to be offloaded based on input information like infrastructure capacities, application requirements, and edge server reputation scores. As described in one of our previous blogs, the decision engine is based on satisfiability modulo theory (SMT), which addresses challenge (C1) by providing formal assurance that relevant edge resource limitations and latency timing constraints are satisfied. SMT relies on input constraints and logic rather than environmental variables, which makes it an environment-agnostic approach, suitable for highly diverse distributed edge environments (C2 challenge).

To address the challenge (C3) of edge failures, we employ a blockchain-based reputation system for estimating the reliability levels of edge servers in the form of long-term performance tracking. However, blockchain-based systems are slowly responsive due to long consensus protocols that conflict with our latency objective, which is our challenge (C4). We employ a hybrid smart contract (HSC) as a reputation state manager to enable a blockchain-based reputation system for latency-sensitive applications. HSC allows off-chain (i.e., outside of blockchain) transactions like fast offloading decisions that require performance while retaining secured on-chain storage (i.e., on the blockchain) of sensitive reputation information against malicious tampering. In this manner, FRESCO bypasses slow blockchain consensus by performing offloading decisions off-chain rather than requiring blockchain consensus for each offloading decision.

FRESCO Edge Offloading Lifecycle Model

To show the interplay between the HSC reputation state manager and SMT offloading decision engine within the FRESCO framework, we introduce the lifecycle model, which is executed as follows. In steps 1a and 1b, the mobile device retrieves the reputation score from HSC and monitors off-chain cluster resources where tasks will be offloaded. Based on the input information, the mobile device calculates offloading decisions and offloads tasks to the off-chain cluster in step 2. Task results are returned to the mobile device in step 3. The mobile device records the performance metric (e.g., response time) and sends it to the HSC on the blockchain for evaluation in step 4. Finally, HSC compares the received performance metric to the timing deadlines and updates the reputation score accordingly. The lifecycle is repeated until the application is terminated. Blockchain consensus is triggered only upon reputation update but not at reputation retrieval, which makes cached reputation score accessible in (near-)real-time.

Experimental Setup and Evaluation Results

We implemented our edge offloading simulator in Python to evaluate our solution against several baselines including MDP, MINLP and SQ approaches, taken from the literature. The infrastructure is simulated based on the OpenCellID dataset that represents radio cell towers, and each location is used as a server location. The workload on the nodes is simulated through the queueing network (blog). For SMT solving, we use Z3 as an SMT solver. We used the Ganache blockchain emulator for the blockchain part and implemented a real-world HSC in the Solidity programming language.

FRESCO outperformed the baselines in decreasing average response time up to 7.86x, and increases battery up to 5.4%, and mixed resource utilization cost, which is expected since we target latency-sensitive applications that prefer faster latency for higher resource utilization cost.

Link to Artifacts

Open access paper: https://arxiv.org/abs/2410.06715

Open source code and data:  https://github.com/jzilic1991/hybrid-edge-blockchain