For an extensive period, Vision Transformers (ViTs) have been deemed unsuitable for attaining robust performance on small-scale datasets, with WideResNet models maintaining dominance in this domain. While WideResNet models have persistently set the state-of-the-art (SOTA) benchmarks for robust accuracy on datasets such as CIFAR-10 and CIFAR-100, this paper challenges the prevailing belief that only WideResNet can excel in this context. We pose the critical question of whether ViTs can surpass the robust accuracy of WideResNet models. Our results provide a resounding affirmative answer. By employing ViT, enhanced with data generated by a diffusion model for adversarial training, we demonstrate that ViTs can indeed outshine WideResNet in terms of robust accuracy. Specifically, under the Infty-norm threat model with epsilon = 8/255, our approach achieves robust accuracies of 74.97% on CIFAR-10 and 44.07% on CIFAR-100, representing improvements of +3.9% and +1.4%, respectively, over the previous SOTA models. Notably, our ViT-B/2 model, with 3 times fewer parameters, surpasses the previously best-performing WRN-70-16. Our achievement opens a new avenue, suggesting that future models employing ViTs or other novel efficient architectures could eventually replace the long-dominant WRN models.

Cloud computing has emerged as an infrastructure in the era of digital economy and has been widely applied in various fields. Virtual Machine(VM) management is the key mechanism in a Cloud Data Center(CDC). A typical VM management system is responsible for VM allocation and VM reallocation, and it is usually designed to optimize specific objectives, especially for the metrics of energy consumption, resource wastage, communication cost, and Service Level Agreement Violations (SLAV). However, it is greatly challenging to optimize these metrics at the same time, and most existing VM management strategies focus on optimizing part of the above four metrics. In this paper, we propose a Holistic Virtual Machine Management (HVMM) strategy to optimize the energy consumption, resource wastage, communication cost, and SLAV simultaneously. First, we define two parameters, the Compatibility and the Performance-to-Power Ratio (PPR), for VM allocation to optimize energy consumption and resource wastage. Then, we propose a reallocation approach based on spectral clustering that can handle dynamic traffic between VMs without a priori knowledge of the traffic between VMs, and it takes a slight expense of resource wastage and energy consumption to reduce communication cost between VMs and ensures low SLAV risk. To evaluate the performance of HVMM, we compared the proposed strategy with the state-of-the-art strategies in various experiments on real-world traces. Compared with the other strategies, the resource wastage of HVMM is reduced by 62%. Simultaneously, the communication cost is reduced by about 26%, the energy consumption is reduced by about 5%, and SLAV is much lower than that of the others.

NVMe SSDs are nowadays widely deployed in various computing platforms due to its high performance and low power consumption, especially in data centers to support modern latency-sensitive applications. NVMe SSDs improve on SATA and SAS interfaced SSDs by providing a large number of device I/O queues at the host side and applications can directly manage the queues to concurrently issue requests to the device. However, the currently deployed request scheduling approach is oblivious to the states of the various device internal components and thus may lead to suboptimal decisions due to various resource contentions at different layers inside the SSD device. In this work, we propose a Conflict Aware Request Scheduling policy named CARS for NVMe SSDs to maximally leverage the rich parallelism available in modern NVMe SSDs. The central idea is to check possible conflicts that a fetched request might be associated with before dispatching that request. If there exists a conflict, it refrains from issuing the request and move to check a request in the next submission queue. In doing so, our scheduler can evenly distribute the requests among the parallel idle components in the flash chips, improving performance. Our evaluations have shown that our scheduler can reduce the slowdown metric by up to 46% relative to the de facto round-robin scheduling policy for a variety of patterned workloads.

Data deduplication, which removes redundant data so that only one copy of duplicate blocks needs to be actually stored, has been implemented in almost all storage appliances, including archival and back-up systems, primary data storage, and SSD devices, to save storage space. However, as time goes and more duplicate blocks have been ingested into the system, the fragmentation problem emerges, that is, logically continuous data blocks of later stored datasets are dispersed in a large storage space and as a result restoring them requires a lot of extra disk accesses, significantly degrading restore performance and garbage collection efficiency. Existing approaches toward the fragmentation problem choose to sacrifice space savings for performance by selectively rewriting trouble-causing duplicate blocks when performing deduplication, even though they have already been stored elsewhere previously. However, rewriting chunks into the system impacts the backup process and reduces deduplication efficiency as many duplicate chunks are allowed in the system. In this work, we propose to deploy flash-based SSDs in the system to overcome the limitations of rewriting algorithms by taking advantage of the high performance provided by SSDs. Specifically, instead of rewriting, we migrate the trouble-causing blocks into an SSD storage in the background when encountering duplicate blocks. The idea is mainly motivated by the following two reasons. First, using a separate migrating process leverages the computing power provided by modern multi-core architecture. Second, typically restores are not performed immediately after backups. Therefore, there is no need to rewrite blocks on the critical path, which affects performance. We augment our proposal to two rewriting schemes and conduct comprehensive evaluations to evaluate its efficacy. Our results show that by provisioning a reasonable amount of SSD, the backup performance and deduplication efficiency can be significantly improved, while slightly increasing the amount of container reads associated with restore operations.