Search

Accurate predictive models for cloud workloads can be helpful in improving task scheduling, capacity planning and preemptive resource conflict resolution, especially in the setting of co-located jobs. Alibaba, one of the leading cloud providers co-locates transient batch tasks and high priority latency sensitive online jobs on the same cluster. In this paper, we consider the problem of using a publicly released dataset by Alibaba to model the batch tasks that are often overlooked compared to online services. The dataset contains the arrivals and resource requirements (CPU, memory, etc.) for both batch and online tasks. Our trained model predicts, with high accuracy, the number of batch tasks that arrive in any 30 minute window, their associated CPU and memory requirements, and their lifetimes. It captures over 94% of arrivals in each 30 minute window within a 95% prediction interval. The F1 scores for the most frequent CPU classes exceed 75%, and our memory and lifetime predictions incur less than 1% test data loss. The prediction accuracy of the lifetime of a batch-task drops when the model uses both CPU and memory information, as opposed to only using memory information.

Performance Isolation in Multi-Tenant Cloud Data Centers (MTCDCs) consists of a set of mechanisms to make sure tenants’ use of resources does not impact other tenants. In this context, traffic shapers and rate limiters are fundamental to addressing the challenges of performance isolation in MTCDCs, which include predictable performance as minimum bandwidth guarantees, tenants-level fairness, and optimal resource utilization. However, the classical linear programming process to find the optimal rates to apply does not scale in terms of computing time, especially with the huge number of nodes, dominated mainly by virtual machines in an MTCDC environment. Motivated by this observation, this paper introduces gPerfIsol, a novel Graph Neural Network (GNN)-based approach designed to find near-optimal rates allocation in near-real-time to ensure performance isolation in MTCDC. gPerfIsol’s key innovation leverages Heterogeneous GNNs to capture MTCDC-specific topological information and demand traffic matrix. Evaluations based on datasets generated through simulation demonstrate the effectiveness of gPerfIsol’s binary classification model with a precision score of 0.964 and a recall score of 0.973. Ultimately, gPerfIso1 offers a promising solution for nearoptimal rate limit allocation for traffic shapers in multi-tenant environments, enhancing performance isolation.

Paper part of Haile's (2023) doctoral thesis Achieving Low Latency and High Throughput over Cellular Internet Connections as manuscript, now published.

Network latency is a critical factor for the perceived quality of experience for many applications. With an increasing focus on interactive and real-time applications, which require reliable and low latency, the ability to continuously and efficiently monitor latency is becoming more important than ever. Always-on passive monitoring of latency can provide continuous latency metrics without injecting any traffic into the network. However, software-based monitoring tools often struggle to keep up with traffic as packet rates increase, especially on contemporary multi-Gbps interfaces. We investigate the feasibility of using eBPF to enable efficient passive network latency monitoring by implementing an evolved Passive Ping (ePPing). Our evaluation shows that ePPing delivers accurate RTT measurements and can handle over 1 Mpps, or correspondingly over 10 Gbps, on a single core, greatly improving on state-of-the-art software based solutions, such as PPing.

Modern information technology (IT) infrastructures are becoming more complex to meet the diverse demands of emerging technology paradigms such as 5G/6G networks, edge, and internet of things (IoT). The intricacy of these infrastructures grows further when hosting containerized workloads as microservices, resulting in the challenge to detect and troubleshoot performance issues, incidents or even outages of critical use cases like industrial automation processes. Thus, fine-grained measurements and associated visualization are essential for operation observability of these IT infrastructures. However, most existing observability tools operate independently without systematically covering the entire data workflow. This paper presents an integrated design for multi-stage observability workflows, denoted as DistributEd obServability frameworK (DESK). The proposed framework aims to improve observability workflows for measurement, collection, fusion, storage, visualization, and notification. As a proof of concept, we deployed the framework in a Kubernetes-based testbed to demonstrate the successful integration of various components and usability of collected observability data. We also conducted a comprehensive study to determine the caused overhead by DESK agents at the reasonably powerful edge node hardware, which shows on average a CPU and memory overhead of around 2.5 % of total available hardware resource.

The promises of multipath transport is to aggregate bandwidth and improve resource utilisation and reliability. We demonstrate in this paper that the way multipath coupled congestion control is defined today RFC6359 leads to a sub-optimal resource utilisation when network paths are mainly disjoint, i.e., they do not share a bottleneck. With growing interest to standardise Multipath QUIC (MPQUIC), we implement the practical shared bottleneck detection (SBD) algorithm from RFC8382 in MPQUIC, namely MPQUIC-SBD. We evaluate MPQUIC-SBD through extensive emulation experiments in the context of video streaming. We show that MPQUIC-SBD is able to correctly detect shared bottlenecks over 90% of the time as the video segments’ size increase depending on the Adaptive Bitrate (ABR) algorithm. In non-shared bottleneck scenarios, MPQUIC-SBD results in video throughput gains of more than 13% compared to MPQUIC, which directly translates into better video quality metrics.

Traffic optimization and smart buffering are fundamental to achieve both great application performance and resource efficiency in data centers with heterogeneous workloads, including incast and elephant traffics. However, general performance models providing insights on how various factors affect traffic performance metrics needed by these management functions are missing. For the special case of incast, the existing models are analytical ones, either tightly coupled with a particular protocol version or specific to certain empirical data. Motivated by this observation, this paper proposes an SDN-enabled machine-learning-based performance modeling approach in data center networks that leverages random forest predictions. Evaluations based on datasets constructed through intensive NS-3 simulations show that we can achieve accurate predictions of incast and elephant performance metrics based on various features. With this performance modeling capability, smart buffering schemes or traffic optimization algorithms could anticipate and efficiently optimize system parameters adjustment to achieve optimal performance continuously in data centers.