While cloud computing providers offer guaranteed allocations for resources such as CPU and memory, they do not offer any guarantees for network resources. The lack of network guarantees prevents tenants from predicting lower bounds on the performance of their applications. The research community has recognized this limitation but, unfortunately, prior solutions have significant limitations: either they are inefficient, because they are not work-conserving, or they are impractical, because they require expensive switch support or congestion-free network cores.

Recent research on memory disaggregation introduces a new architectural building block — the memory blade — as a cost-effective approach for memory capacity expansion and sharing for an ensemble of blade servers. Memory blades augment blade servers' local memory capacity with a second-level (remote) memory that can be dynamically apportioned among blades in response to changing capacity demand, albeit at a higher access latency. In this paper, we build on the prior research to explore the software and systems implications of disaggregated memory. We develop a software-based prototype by extending the Xen hypervisor to emulate a disaggregated memory design wherein remote pages are swapped into local memory on-demand upon access. Our prototyping effort reveals that low-latency remote memory calls for a different regime of replacement policies than conventional disk paging, favoring minimal hypervisor overhead even at the cost of using less sophisticated replacement policies. Second, we demonstrate the synergy between disaggregated memory and content-based page sharing. By allowing content to be shared both within and across blades (in local and remote memory, respectively), we find that their combination provides greater workload consolidation opportunity and performance-per-dollar than either technique alone. Finally, we explore a realistic deployment scenario in which disaggregated memory is used to reduce the scaling cost of a memcached system. We show that disaggregated memory can provide a 50% improvement in performance-per-dollar relative to conventional scale-out.

Software Defined Networking, or SDN, based networks are being deployed not only in testbed networks, but also in production networks. Although fault-tolerance is one of the most desirable properties in production networks, there are not much study in providing fault-tolerance to SDN-based networks. The goal of this work is to develop a fault tolerant SDN architecture that can rapidly recover from faults and scale to large network sizes. This paper presents CORONET, a SDN fault-tolerant system that recovers from multiple link failures in the data plane. We describe a prototype implementation based on NOX that demonstrates fault recovery for emulated topologies using Mininet. We also discuss possible extensions to handle control plane and controller faults.

Virtualization has fundamentally changed the data center network. The last hop of the network is no longer handled by a physical network switch, but rather is typically performed in software inside the server to switch among virtual machines hosted by that server.

Most datacenter networks are based on specialized edge-core topologies, which are costly to build, difficult to maintain and consume too much power. We propose enhancements to layer-two (L2) Ethernet switches to enable multipath L2 routing in scalable datacenters. This replaces an expensive router with commodity switches. Our hash-based routing approach reuses and minimally extends hardware structures in high-volume switches, while exposing a powerful network management inter-face for multipath load balancing, QoS differentiation, and resilience to faults. Simulation results demonstrate near-optimal load balancing for uniform and non-uniform traffic patterns, and effective management of large datacenter networks independent of the number of traffic flows.

Most datacenter networks are based on specialized edge-core topologies, which are costly to build, difficult to maintain and consume too much power. We propose enhancements to layer-two (L2) Ethernet switches to enable multipath L2 routing in scalable datacenters. Our hash-based routing approach reuses and minimally extends hardware structures in high-volume switches, while exposing a powerful network management interface for multipath load balancing, QoS differentiation, and resilience to faults.

This paper presents mechanisms and optimizations to reduce the overhead of network interface virtualization when using the driver domain I/O virtualization model. The driver domain model provides benefits such as support for legacy device drivers and fault isolation. However, the processing overheads incurred in the driver domain to achieve these benefits limit overall I/O performance. This paper demonstrates the effectiveness of two approaches to reduce driver domain overheads. First, Xen is modified to support multi-queue network interfaces to eliminate the software overheads of packet demultiplexing and copying. Second, a grant reuse mechanism is developed to reduce memory protection overheads. These mechanisms shift the bottleneck from the driver domain to the guest domains, improving scalability and enabling significantly higher data rates. This paper also presents and evaluates a series of optimizations that substantially reduce the I/O virtualization overheads in the guest domain. In combination, these mechanisms and optimizations increase the maximum throughput achieved by guest domains from 2.9Gb/s to full 10 Gigabit Ethernet link rates.

We present a new distributed checkpoint-restart mechanism, Cruz, that works without requiring application, library, or base kernel modifications. This mechanism provides comprehensive support for checkpointing and restoring application state, both at user level and within the OS. Our implementation builds on Zap, a process migration mechanism, implemented as a Linux kernel module, which operates by interposing a thin layer between applications and the OS. In particular, we enable support for networked applications by adding migratable IP and MAC addresses, and checkpoint-restart of socket buffer state, socket options, and TCP state. We leverage this capability to devise a novel method for coordinated checkpoint-restart that is simpler than prior approaches. For instance, it eliminates the need to flush communication channels by exploiting the packet re-transmission behavior of TCP and existing OS support for packet filtering. Our experiments show that the overhead of coordinating checkpoint-restart is negligible, demonstrating the scalability of this approach.

Virtual Machine (VM) environments (e.g., VMware and Xen) are experiencing a resurgence of interest for diverse uses including server consolidation and shared hosting. An application's performance in a virtual machine environment can differ markedly from its performance in a non-virtualized environment because of interactions with the underlying virtual machine monitor and other virtual machines. However, few tools are currently available to help debug performance problems in virtual machine environments.In this paper, we present Xenoprof, a system-wide statistical profiling toolkit implemented for the Xen virtual machine environment. The toolkit enables coordinated profiling of multiple VMs in a system to obtain the distribution of hardware events such as clock cycles and cache and TLB misses. The toolkit will facilitate a better understanding of performance characteristics of Xen's mechanisms allowing the community to optimize the Xen implementation.We use our toolkit to analyze performance overheads incurred by networking applications running in Xen VMs. We focus on networking applications since virtualizing network I/O devices is relatively expensive. Our experimental results quantify Xen's performance overheads for network I/O device virtualization in uni- and multi-processor systems. With certain Xen configurations, networking workloads in the Xen environment can suffer significant performance degradation. Our results identify the main sources of this overhead which should be the focus of Xen optimization efforts. We also show how our profiling toolkit was used to uncover and resolve performance bugs that we encountered in our experiments which caused unexpected application behavior.

Large-scale systems experience frequent failures which can result in unacceptably high service downtime or application execution time. To meet performance and availability requirements, the user must perform a complex design task including the selection and configuration of hardware and software components and mechanisms for handling failures. We believe users should be relieved of this burden by automating the design process in order to generate cost-effective solutions from high-level application requirements. In this paper, we present Aved, a proof of concept design automation engine which is a first step toward this goal. We describe how infrastructure choices, application models, and user requirements are represented with Aved to automate design space search and reason about design alternatives. We additionally present examples to illustrate how Aved can generate a complete picture of the cost-availability and cost-performance tradeoffs for the infrastructure design.

Infiniband system area networks (SANs) which use link-level flow control experience congestion spreading, where one bottleneck link causes traffic to block throughout the network. In this paper, we propose an end-to-end congestion control scheme that avoids congestion spreading, delivers high throughput, and prevents flow starvation. It couples a simple switch-based ECN packet marking mechanism appropriate for typical SAN switches with small input buffers, together with a source response mechanism that uses rate control combined with a window limit. The classic fairness convergence requirement for source response functions assumes network feedback is synchronous. We relax the classic requirement by exploiting the asynchronous behavior of packet marking. Our experimental results demonstrate that compared to conventional approaches, our proposed marking mechanism improves fairness. Moreover, rate increase functions possible under the relaxed requirement reclaim available bandwidth aggressively and improve throughput in both static and dynamic traffic scenarios.

Internet services experience frequent changes in demand, resource characteristics and service requirements. A service can meet its requirements with minimal cost only if the allocation of its distributed resources can be adaptively controlled. The design of adaptive control schemes requires a solid understanding of the dynamic characteristics of the distributed service, demand and resources. This study examines the dynamic properties of distributed demand and differs from prior demand characterization work by focusing on dynamic variations across clients, time and region which are crucial in adaptively allocating distributed resources. Our analysis of the demand for the 1998 World Cup web site finds that the dynamic behavior of a small subset of clients is representative of the entire demand, the churn in the active set of clients from day to day is relatively small and that regional demand shows significant, and predictable, variations in the hourly scales. These results will provide guidance for the design of adaptive policies.

InfiniBand System Area Networks using link-level flow control can experience congestion spreading, where one bottleneck link causes traffic to block throughout the network. End-to-end congestion control using Explicit Congestion Notification (ECN) packet marking has the potential to solve this problem. In this paper, we develop and evaluate ECN mechanisms for switches with input-buffered configuration (typical for InfiniBand), as opposed to traditional techniques that target output-buffered switches. Our mechanisms vary in two dimensions: the condition that triggers a marking event, and the set of packets marked at such an event. The experimental results show that our approaches avoid congestion spreading while preserving high network throughput. In addition, selective marking of packets at an input buffers on each marking event provides better fairness than a naive mechanism that marks only packets in a full input buffer. Finally, schemes that use state at both input and output ports to trigger marking events can improve fairness over purely input-triggered marking schemes.

We compare performance of a multimedia storage server based on a random data allocation layout and block replication with traditional data striping techniques. Data striping techniques in multimedia servers are often designed for restricted workloads, e.g. sequential access patterns with CBR (constant bit rate) requirements. On the other hand, a system based on random data allocation can support virtually any type of multimedia application, including VBR (variable bit rate) video or audio, and interactive applications with unpredictable access patterns, such as 3D interactive virtual worlds, interactive scientific visualizations, etc. Surprisingly, our results show that system performance with random data allocation is competitive and sometimes even outperforms traditional data striping techniques, for the workloads for which data striping is designed to work best; i.e. streams with sequential access patterns and CBR requirements. Due to its superiority in supporting general workloads and competitive system performance, we believe that random data allocation will be the scheme of choice for next generation multimedia servers.

HO is a mtdtimedia object server which manages a set of parfld disks and supports red-time data d&very tith statisticrd dday guarantees.MO uses random data aRocation on diik< combiied with partial rephcation to achieve load balance and high performance b this paper we amdyze the perform~ce of NO when the set, of disks used to store data blocks is not homogeneous, having both tierent bandwidths and Merent storage capacities.The basic problem to be addr~ tems, Intel Corp., md NSF -t ~-952717S.t.USO i~ith Universityof Sao Pmslq h= resear&k ptiidy SUP Fotied by a f~oIrshlF from C~q @til) Perrnksionlo ruake digitdorhardcopi= of d] orpanof Ihii urorkfor personalor classroomuse is granted withoutfee provided that copies are not made or distributed for profitor commercial advantage, andthat copiesbear!hk noticeand~hefull citationon thefirst page.To copyolhewise,to republishto postonsewers or lo redistributeIO lists, requires prior specific permission andlora fee.

We describe design and implementation of the RIO (randomized I/O) multimedia object server which manages a set of parallel disks and supports real-time throughput and statistical delay guarantees. This storage subsystem was implemented as part of a multimedia information server supporting multiple concurrent applications such as video on demand and 3D interactive virtual worlds. We discuss principal issues and innovations involved in design and implementation of RIO storage system, presenting experimental performance results measured on a prototype implementation. A multimedia data server must be ready to handle a variety of real-time object types (video, audio, 3D interactive virtual worlds, etc.) along with a nonreal-time workload component. Achieving simultaneously (1) high utilization and (2) low latency with high degree of certainty is the challenge. Our prototype system provides a statistical guarantee of quality of service. Our experimental results show that it is possible to achieve a very small probability of exceeding a deadline (less than 10−6), with relatively high disk utilization (70–99%, in terms of fraction of the maximum disk capacity), together with relatively small delay bound (on the order of 0.5–1.5 s), using contemporary disks. Results were achieved using random allocation of disk blocks and replication in conjunction with on-line load balancing. © 1998 John Wiley & Sons, Inc.