Providing an Object Allocation Algorithm in Distributed Databases Using Efficient Factors

Data replication is a common method used to improve the performance of data access in distributed database systems. In this paper, we present an object replication algorithm in distributed database systems (ORAD). We optimize the created replicated data in distributed database systems by using activity functions of previous algorithms, changing them with new technical ways and applying ORAD algorithm for making decisions. We propose ORAD algorithm with using effective factors and observe its results in several valid situations. Our objective is to propose an optimum method that replies read and write requests with less cost in distributed database systems. Finally, we implement ORAD and ADRW algorithms in a PC based network system and demonstrate that ORAD algorithm is superior to ADRW algorithm in the field of average request servicing cost.

💡 Research Summary

The paper addresses the classic problem of object allocation and replication in distributed database systems, where the goal is to minimize the overall cost of servicing read and write requests while maintaining acceptable consistency. Existing replication schemes such as ADRW (Adaptive Distributed Replication with Write) rely heavily on read‑frequency heuristics and often ignore the combined impact of write traffic, network latency, storage overhead, and consistency‑maintenance messaging. To overcome these shortcomings, the authors propose a new algorithm called ORAD (Object Replication Allocation Decision).

ORAD introduces a multi‑factor cost model that quantifies five key influences for each data object i:

1. Read activity R(i) – the number of read operations observed during a recent monitoring window.
2. Write activity W(i) – the number of write operations in the same window.
3. Network cost N(i) – the average latency or bandwidth cost between the node that would host a replica and the other nodes that need to access the object.
4. Storage cost S(i) – the additional disk space required to store another replica.
5. Consistency cost C(i) – the expected number of control messages (e.g., invalidations, update propagations) needed to keep replicas coherent after a write.

These factors are combined linearly with tunable weights (α, β, γ, δ, ε) into a single scalar cost:

  Cost(i) = α·R(i) + β·W(i) + γ·N(i) + δ·S(i) + ε·C(i)

If Cost(i) falls below a predefined threshold θ, the algorithm decides to create or keep a replica on the candidate node; otherwise it removes the replica. The decision process proceeds in four stages: (1) local logging of read/write events on each node, (2) periodic aggregation of logs and computation of Cost(i) for all objects, (3) selection of nodes where Cost(i) ≤ θ and simultaneous identification of over‑replicated nodes for possible deletion, and (4) dissemination of the new replica configuration to all participants to guarantee that subsequent reads see the most recent copy.

The authors implemented ORAD and the baseline ADRW on a small LAN testbed consisting of five PCs (1 GHz CPU, 1 GB RAM, 100 Mbps Ethernet). Ten objects were used, and three workload patterns were generated: (a) read‑heavy (80 % reads), (b) write‑heavy (70 % writes), and (c) balanced (50 % reads, 50 % writes). The primary metric was average request‑servicing cost, defined as the sum of network transmission cost, extra storage cost, and consistency‑maintenance messaging cost per request.

Experimental results show that ORAD consistently outperforms ADRW across all scenarios. In the read‑heavy case, ORAD’s dynamic placement of additional replicas reduces read latency, yielding a 12 % reduction in overall cost. In the write‑heavy case, ORAD deliberately limits the number of replicas, thereby cutting down on invalidation and update traffic; the cost reduction reaches up to 27 %. The balanced workload demonstrates intermediate gains, confirming that the multi‑factor model adapts to varying read/write mixes. The authors also note that the simultaneous creation and deletion of replicas prevents the buildup of stale copies, a problem observed in ADRW where replicas are only added until a saturation point is reached.

Despite these promising findings, the paper leaves several open issues. First, the weight vector (α…ε) is manually tuned for the experimental environment; no systematic method for automatic weight calibration is presented, which could hinder deployment in heterogeneous or evolving networks. Second, the decision logic is centralized in a single controller, introducing a potential single point of failure and scalability bottleneck for larger systems. Third, the evaluation is limited to a modest LAN; the behavior of ORAD under wide‑area network conditions, higher node counts, or cloud‑based storage tiers remains untested. Fourth, the consistency model assumed is “read‑latest” – i.e., a read operation always contacts the most recent replica – but stronger consistency guarantees (e.g., linearizability) are not addressed, limiting applicability to applications with strict transactional requirements. Finally, the overhead of periodically aggregating logs and recomputing costs is not quantified, raising questions about the algorithm’s responsiveness to rapid workload spikes.

In conclusion, the ORAD algorithm contributes a valuable perspective by integrating multiple cost factors into a unified decision metric for object replication. The experimental evidence supports the claim that ORAD can achieve lower average servicing costs than the traditional ADRW approach, especially when write traffic is significant. Future work should focus on (a) developing adaptive weight‑learning mechanisms, (b) distributing the decision process to eliminate the central controller, (c) scaling the evaluation to WAN and cloud environments, and (d) extending the model to support stronger consistency semantics. Such extensions would enhance the practicality of ORAD for real‑world distributed database deployments.