Search

We design and implement LEGOStore, an erasure coding (EC) based linearizable data store over geo-distributed public cloud data centers (DCs). For such a data store, the confluence of the following factors opens up opportunities for EC to be latency-competitive with replication: (a) the necessity of communicating with remote DCs to tolerate entire DC failures and implement linearizability; and (b) the emergence of DCs near most large population centers. LEGOStore employs an optimization framework that, for a given object, carefully chooses among replication and EC, as well as among various DC placements to minimize overall costs. To handle workload dynamism, LEGOStore employs a novel agile reconfiguration protocol. Our evaluation using a LEGOStore prototype spanning 9 Google Cloud Platform DCs demonstrates the efficacy of our ideas. We observe cost savings ranging from moderate (5-20\%) to significant (60\%) over baselines representing the state of the art while meeting tail latency SLOs. Our reconfiguration protocol is able to transition key placements in 3 to 4 inter-DC RTTs ($<$ 1s in our experiments), allowing for agile adaptation to dynamic conditions., Comment: Extended version of paper to appear in PVLDB 2022

We consider a content-caching system thatis shared by a number of proxies. The cache could belocated in an edge-cloud datacenter and the proxies couldeach serve a large population of mobile end-users. Eachproxy operates its own LRU-list of a certain capacity inthe shared cache. The length of objects simultaneouslyappearing in plural LRU-lists is equally divided amongthem,i.e., object sharing among the LRUs. We provide a "working-set" approximation for this system to quicklyestimate the cache-hit probabilities under such objectsharing, which can be used to facilitate admission control.Also, a way to reduce ripple evictions,i.e.,setrequestoverhead, is suggested. We give numerical results for ourMemCacheD with Object Sharing (MCD-OS) prototype.

Using tiny, equal-sized tasks (Homogeneous microTasking, HomT) has long been regarded an effective way of load balancing in parallel computing systems. When combined with nodes pulling in work upon becoming idle, HomT has the desirable property of automatically adapting its load distribution to the processing capacities of participating nodes - more powerful nodes finish their work sooner and, therefore, pull in additional work faster. As a result, HomT is deemed especially desirable in settings with heterogeneous (and possibly possessing dynamically changing) processing capacities. However, HomT does have additional scheduling and I/O overheads that might make this load balancing scheme costly in some scenarios. In this paper, we first analyze these advantages and disadvantages of HomT. We then propose an alternative load balancing scheme - Heterogeneous MacroTasking (HeMT) - wherein workload is intentionally partitioned according to nodes' processing capacity. Our goal is to study when HeMT is able to overcome the performance disadvantages of HomT. We implement a prototype of HeMT within the Apache Spark application framework with complementary enhancements to the Apache Mesos cluster manager. Spark's built-in scheduler, when parameterized appropriately, implements HomT. Our experimental results show that HeMT out-performs HomT when accurate workload-specific estimates of nodes' processing capacities are learned. As representative results, Spark with HeMT offers about 10% better average completion times for realistic data processing workloads over the default system.

In the following, we present example illustrative and experimental results comparing fair schedulers allocating resources from multiple servers to distributed application frameworks. Resources are allocated so that at least one resource is exhausted in every server. Schedulers considered include DRF (DRFH) and Best-Fit DRF (BF-DRF), TSF, and PS-DSF. We also consider server selection under Randomized Round Robin (RRR) and based on their residual (unreserved) resources. In the following, we consider cases with frameworks of equal priority and without server-preference constraints. We first give typical results of a illustrative numerical study and then give typical results of a study involving Spark workloads on Mesos which we have modified and open-sourced to prototype different schedulers.

Efficient and fair allocation of multiple types of resources is a crucial objective in a cloud/distributed computing cluster. Users may have diverse resource needs. Furthermore, diversity in server properties/ capabilities may mean that only a subset of servers may be usable by a given user. In platforms with such heterogeneity, we identify important limitations in existing multi-resource fair allocation mechanisms, notably Dominant Resource Fairness (DRF) and its follow-up work. To overcome such limitations, we propose a new server-based approach; each server allocates resources by maximizing a per-server utility function. We propose a specific class of utility functions which, when appropriately parameterized, adjusts the trade-off between efficiency and fairness, and captures a variety of fairness measures (such as our recently proposed Per-Server Dominant Share Fairness). We establish conditions for the proposed mechanism to satisfy certain properties that are generally deemed desirable, e.g., envy-freeness, sharing incentive, bottleneck fairness, and Pareto optimality. To implement our resource allocation mechanism, we develop an iterative algorithm which is shown to be globally convergent. Finally, we show how the proposed mechanism could be implemented in a distributed fashion. We carry out extensive trace-driven simulations to show the enhanced performance of our proposed mechanism over the existing ones., Comment: Technical Report, 20 pages, 14 figures

We first consider the static problem of allocating resources to ( i.e. , scheduling) multiple distributed application framework s, possibly with different priorities and server preferences , in a private cloud with heterogeneous servers. Several fai r scheduling mechanisms have been proposed for this purpose. We extend pr ior results on max-min and proportional fair scheduling to t his constrained multiresource and multiserver case for generi c fair scheduling criteria. The task efficiencies (a metric r elated to proportional fairness) of max-min fair allocations found b y progressive filling are compared by illustrative examples . They show that "server specific" fairness criteria and those that are b ased on residual (unreserved) resources are more efficient.

Users of cloud computing platforms pose different types of demands for multiple resources on servers (physical or virtual machines). Besides differences in their resource capacities, servers may be additionally heterogeneous in their ability to service users - certain users' tasks may only be serviced by a subset of the servers. We identify important shortcomings in existing multi-resource fair allocation mechanisms - Dominant Resource Fairness (DRF) and its follow up work - when used in such environments. We develop a new fair allocation mechanism called Per-Server Dominant-Share Fairness (PS-DSF) which we show offers all desirable sharing properties that DRF is able to offer in the case of a single "resource pool" (i.e., if the resources of all servers were pooled together into one hypothetical server). We evaluate the performance of PS-DSF through simulations. Our evaluation shows the enhanced efficiency of PS-DSF compared to the existing allocation mechanisms. We argue how our proposed allocation mechanism is applicable in cloud computing networks and especially large scale data-centers., Comment: 12 pages, 7 figures, technical report

We formulate optimization problems to study how data centers might modulate their power demands for cost-effective operation taking into account three key complex features exhibited by real-world electricity pricing schemes: (i) time-varying prices (e.g., time-of-day pricing, spot pricing, or higher energy prices during coincident peaks) and (ii) separate charge for peak power consumption. Our focus is on demand modulation at the granularity of an entire data center or a large part of it. For computational tractability reasons, we work with a fluid model for power demands which we imagine can be modulated using two abstract knobs of demand dropping and demand delaying (each with its associated penalties or costs). Given many data center workloads and electric prices can be effectively predicted using statistical modeling techniques, we devise a stochastic dynamic program (SDP) that can leverage such predictive models. Since the SDP can be computationally infeasible in many real platforms, we devise approximations for it. We also devise fully online algorithms that might be useful for scenarios with poor power demand or utility price predictability. For one of our online algorithms, we prove a competitive ratio of 2-1/n. Finally, using empirical evaluation with both real-world and synthetic power demands and real-world prices, we demonstrate the efficacy of our techniques. As two salient empirically-gained insights: (i) demand delaying is more effective than demand dropping regarding to peak shaving (e.g., 10.74% cost saving with only delaying vs. 1.45% with only dropping for Google workload) and (ii) workloads tend to have different cost saving potential under various electricity tariffs (e.g., 16.97% cost saving under peak-based tariff vs. 1.55% under time-varying pricing tariff for Facebook workload).

Since the electricity bill of a data center constitutes a significant portion of its overall operational costs, reducing this has become important. We investigate cost reduction opportunities that arise by the use of uninterrupted power supply (UPS) units as energy storage devices. This represents a deviation from the usual use of these devices as mere transitional fail-over mechanisms between utility and captive sources such as diesel generators. We consider the problem of opportunistically using these devices to reduce the time average electric utility bill in a data center. Using the technique of Lyapunov optimization, we develop an online control algorithm that can optimally exploit these devices to minimize the time average cost. This algorithm operates without any knowledge of the statistics of the workload or electricity cost processes, making it attractive in the presence of workload and pricing uncertainties. An interesting feature of our algorithm is that its deviation from optimality reduces as the storage capacity is increased. Our work opens up a new area in data center power management., Comment: Full version of Sigmetrics 2011 paper