Abstract With the growing scale and complexity of high-performance computing systems, ensuring reliability through robust fault diagnosis becomes increasingly critical. System-level diagnosis plays a key role in identifying faulty processors and maintaining system stability of multiprocessor systems. However, traditional diagnosability, as a global reliability metric for multiprocessor systems, overlooks local diagnostic capability, topological criticality, and fault distribution. In order to better capture the local characteristics of a system around a given node, this work proposes a novel fault diagnosis strategy, called cyclic local diagnosability, where the cyclic fault pattern requires that at least two components contain cycles. We propose some characterizations of cyclic local diagnosability of interconnection networks under PMC and MM* models. As applications, we determine the cyclic local diagnosabilities of data center network DCell ($D_{k,n}$), $(n,k)$-star graph ($S_{n,k}$) and $(n,k)$-bubble-sort graph ($B_{n,k}$) under PMC and MM* models. Finally, we show the superiority of the cyclic local diagnosability through comparison with other conditional diagnosabilities.

The rapid growth of cloud computing, big data processing, and high-performance applications has led to the urgent demand for fault-tolerant and scalable data center networks. In this paper, we propose a novel data center network, DCube, based on the$n$-dimensional dual-cube, which outperforms other newly proposed data center networks, such as HSDC, CSDC, and HHCube, in both bisection width and scalability. We first propose the shortest path algorithm within the logical graph${\mathcal {DC}}_{n}$of DCube and show its diameter and bisection width as$4n + 1$and$2^{2n-3}$, respectively. Next, we determine the$g$-good-neighbor connectivity and$g$-extra connectivity of${\mathcal {DC}}_{n}$, as well as the node and edge connectivity of${\mathcal {DC}}_{n}$. We then explore the$g$-good-neighbor diagnosability and$g$-extra diagnosability of${\mathcal {DC}}_{n}$based on the derived$g$-good-neighbor connectivity and$g$-extra connectivity. Finally, we propose a N-E-FD algorithm for fault identification with time complexity of$O(n^{2} 2^{2n-1})$, and validate its effectiveness and accuracy through simulations. The results demonstrate that DCube provides a robust, scalable, and efficient choice for the data center architectures.

The rapid advancement of semiconductor technology has enabled the development of large-scale multiprocessor systems, which are crucial for high-performance computing systems, data centers, and cloud infrastructures. However, as these systems grow in complexity and scale, the assessment of reliability becomes an urgent issue that needs to be solved, which calls for effective fault diagnosis to detect failures and maintain system performance. To this end, the concept of$(t,k)$-diagnosis was introduced, which detects all failing nodes when their number is at most$k$, and otherwise identifies at least$k$failing nodes per iteration as long as the total number does not exceed$t$. While traditional fault diagnosis strategies are effective, they often encounter the restriction of objective circumstances. For instance, it is improbable that all nodes adjacent to a particular node fail simultaneously. To enhance the efficiency and accuracy of fault diagnosis, this work introduces the$g$-good-neighbor conditional$(t,k)$-diagnosis. It ensures each node owns no fewer than$g$neighbors that are fault-free to match actual environmental requirements. For a general multiprocessor system modeled by$G$, let$\kappa _{g}(G)$be the$g$-good-neighbor connectivity, and define$\delta$as the minimum degree and$\Delta$as the maximum degree of$G$. Under the PMC model, we not only propose two conditional$(t,k)$-diagnosis algorithms, but also prove that$G$is$g$-good-neighbor conditional$(\frac{g|V|-p}{\Delta +g-1}, \min \lbrace p, k_{g}(G)\rbrace)$-diagnosable when fathomed components arise, where$g\leq \lfloor \frac{\Delta +1}{2}\rfloor$and$p\geq 1$, while in the absence of fathomed components,$G$is$g$-good-neighbor conditional$(\frac{g|V| + |B| - 2}{\Delta +g-1}, \kappa _{g}(G))$-diagnosable, where$g\leq \min \lbrace \delta -2,\lfloor \frac{\Delta }{2}\rfloor \rbrace$,$\lambda = (g+1)|B|$,$|B| < \frac{2\lambda (|B| - 1)+(\delta \lambda - \hat{I}(\lambda))|V|}{(\Delta +\delta) \lambda - \hat{I}(\lambda)}$, and$\hat{I}(\lambda)$approximates the number of directed edges among$\lambda$nodes. Experimental results show that conditional$(t,k)$-diagnosis algorithms achieve perfect fault identification with low runtime and good scalability.