Current Data Center Research Articles

The RedFish API and vSphere Hypervisor API: A Unified Framework for Policy-Based Server Monitoring

Integrating remote monitoring systems is crucial in the ever-changing field of data center management to enhance performance and guarantee reliability. This paper outlines a comprehensive strategy for monitoring remote servers by utilizing agents that establish connections to the RedFish API (Application Programming Interface) and vSphere hypervisor API. Our solution uses the RedFish standard to provide secure and standardized management of hardware components in diverse server environments. This improves interoperability and scalability. Simultaneously, the vSphere agent enables monitoring and hardware administration in vSphere-based virtualized environments, offering crucial insights into the state of the underlying hardware. This system, which employs two agents, simplifies the management of servers and seamlessly integrates with current data center infrastructures, enhancing efficiency. The policy-based alerting system built on top of these agents offers many capabilities based on both agents leveraging their alerting systems. This, in turn, can improve the capabilities of next-generation data centers.

12.1 terabit/second data center interconnects using O-band coherent transmission with QD-MLL frequency combs

Most current Data Center Interconnects (DCI) use intensity modulation direct detection (IMDD) configurations due to their low complexity and cost. However, significant scaling challenges allow coherent solutions to become contenders in these short reach applications. We present an O-band coherent optical fiber transmission system based on Quantum Dot—Mode Locked Lasers (QD-MLLs) using two independent free-running comb lasers, one each for the carrier and the Local Oscillator (LO). Using a comb-to-comb configuration, we demonstrate a 10 km single mode fiber, O-band, coherent, heterodyne, 12.1 Tbps system operating at 0.47 Tbps/λ using 26 λs. We used fewer comb lines (26 λs), faster symbol rate (56 GBaud) and higher constellation cardinality (32 QAM) relative to the highest capacity C-band systems reported to date. Through design, analysis, and experimentation, we quantify the optimum comb line spacing for this use case. We compare potential configurations for increasing data center interconnect capacities whilst reducing power consumption, complexity, and cost.

Feasibility investigation on a novel data center cooling system based on cross-seasonal soil cool storage in severe cold area

In response to excessive energy consumption and high carbon emissions associated with the current data center cooling system, the advancement in the utilization of instant free cooling has progressively halted. A novel data center cooling system based on cross-seasonal soil cold storage is proposed，which makes full use of the cold stored in the soil across the seasons and air cold source for realizing the year-round free cooling. A dynamical simulation platform, which describes its operation characteristic, is established to analyze the feasibility. Results indicate that in the typical severe cold area of Harbin, the soil cooling during summer is sufficient to meet the cooling requirements of the data center. The maximum outlet temperature of the ground heat exchanger is only 8.9 °C when releasing cold. The cold energy stored across seasons in the soil can be effectively utilized, achieving a utilization rate of 98.04 %. In Harbin, Changchun, and Shenyang, the cooling capacity stored across seasons accounts for 31.4 %, 32.2 %, and 36.9 % of the total annual cooling capacity of the system, respectively. The annual average coefficient of performance for the system is 13.63, 12.52, and 10.44, respectively. Compared to the cooling system that does not utilize cross-seasonal soil cool storage but achieves efficient use of instant natural cold sources, these values represent an increase of 4.92, 4.02 and 2.62, respectively. Overall, this system demonstrates excellent applicability in severe cold area.

Coding, Decoding and Retrieving a Message Using DNA: An Experience from a Brazilian Center Research on DNA Data Storage.

DNA data storage based on synthetic oligonucleotides is a major attraction due to the possibility of storage over long periods. Nowadays, the quantity of data generated has been growing exponentially, and the storage capacity needs to keep pace with the growth caused by new technologies and globalization. Since DNA can hold a large amount of information with a high density and remains stable for hundreds of years, this technology offers a solution for current long-term data centers by reducing energy consumption and physical storage space. Currently, research institutes, technology companies, and universities are making significant efforts to meet the growing need for data storage. DNA data storage is a promising field, especially with the advancement of sequencing techniques and equipment, which now make it possible to read genomes (i.e., to retrieve the information) and process this data easily. To overcome the challenges associated with developing new technologies for DNA data storage, a message encoding and decoding exercise was conducted at a Brazilian research center. The exercise performed consisted of synthesizing oligonucleotides by the phosphoramidite route. An encoded message, using a coding scheme that adheres to DNA sequence constraints, was synthesized. After synthesis, the oligonucleotide was sequenced and decoded, and the information was fully recovered.

DemePro: DEcouple packet Marking from Enqueuing for multiple services with PROactive congestion control

Most of current Data Center Network (DCN) protocols leverage Explicit Congestion Notification (ECN) for congestion control to maintain low latency and high throughput. However, the majority of them only consider single-queue scenario in each switch port, making their performance inferior in multiple-service multiple-queue scenario. To this end, we propose DemePro, a DCN scheme for multiple-service multiple-queue scenario. In face of congestion signal, DemePro also leverages ECN for congestion notification, while decouples packet marking from enqueuing, in order to ensure fairness among multiple services. We also question the effectiveness of the congestion signal derived from the single threshold for port buffer queue length, via a set of experiments. Then, DemePro utilizes multiple thresholds for proactive congestion control, and packets are encapsulated to carry congestion extent information, in order to notify TCP senders for precise and fine-grained congestion control. Experiments show that DemePro has better performance than MQ-ECN [1] which is currently the best DCN scheme for multiple-service multiple-queue scenario, and DemePro can also well guarantee the fairness.

Reinforcement learning based task scheduling for environmentally sustainable federated cloud computing

The significant energy consumption within data centers is an essential contributor to global energy consumption and carbon emissions. Therefore, reducing energy consumption and carbon emissions in data centers plays a crucial role in sustainable development. Traditional cloud computing has reached a bottleneck, primarily due to high energy consumption. The emerging federated cloud approach can reduce the energy consumption and carbon emissions of cloud data centers by leveraging the geographical differences of multiple cloud data centers in a federated cloud. In this paper, we propose Eco-friendly Reinforcement Learning in Federated Cloud (ERLFC), a framework that uses reinforcement learning for task scheduling in a federated cloud environment. ERLFC aims to intelligently consider the state of each data center and effectively harness the variations in energy and carbon emission ratios across geographically distributed cloud data centers in the federated cloud. We build ERLFC using Actor-Critic algorithm, which select the appropriate data center to assign a task based on various factors such as energy consumption, cooling method, waiting time of the task, energy type, emission ratio, and total energy consumption of the current cloud data center and the details of the next task. To demonstrate the effectiveness of ERLFC, we conducted simulations based on real-world task execution data, and the results show that ERLFC can effectively reduce energy consumption and emissions during task execution. In comparison to Round Robin, Random, SO, and GJO algorithms, ERLFC achieves respective reductions of 1.09, 1.08, 1.21, and 1.26 times in terms of energy saving and emission reduction.

DAG-based workflows scheduling using Actor–Critic Deep Reinforcement Learning

High-Performance Computing (HPC) is essential to support the advance in multiple research and industrial fields. Despite the recent growth in processing and networking power, the HPC Data Centers (DCs) are finite, and should be carefully managed to host multiple jobs. The scheduling of tasks (composing a job) is a crucial and complex task, once the reflexes of the scheduler’s decisions are perceptible both for users (e.g., slowdown) and for infrastructure administrators (e.g., use of resources and queue length). In fact, the process of scheduling workflows atop a DC can be modeled as a graph mapping problem. While an undirected graph is used to represent the DC, a Directed Acyclic Graph (DAG) is used to express the tasks dependencies. Each vertex and edge from both graphs can have weights associated with them, denoting the residual capacities for DC resources, as well as computing and networking demands for workflows. Motivated by the combinatorial explosion of the aforementioned scheduling problem, the integration of Machine Learning (ML) for generating or improving scheduling policies is a reality, however the proposals in the specialized literature opt, mostly, for using simplified models to reduce the search space or are trained to specific scenarios, which leads to policies that eventually fall short of real DCs expectations. Given this challenge, this work applies Actor–Critic (AC) Reinforcement Learning (RL) to schedule DAG-based workflows. Instead of proposing a new policy, the AC RL is used to select the appropriated scheduling policy from a pool of consolidated algorithms, guided by the DAGs workload and DC usage. The AC RL-based scheduler analyzes the DAGs queue and the DC status to define which algorithms are better suited to improve the overall performance indicators in each scenario instance. The simulation protocol comprises multiple analysis with distinct workload configurations, number of jobs, queue ordering polices and strategies to select the target DC servers. The results demonstrated that the AC RL selects the scheduling policy which fits the current workload and DC status.

Using Local Cache Coherence for Disaggregated Memory Systems

Disaggregated memory provides many cost savings and resource provisioning benefits for current datacenters, but software systems enabling disaggregated memory access result in high performance penalties. These systems require intrusive code changes to port applications for disaggregated memory or employ slow virtual memory mechanisms to avoid code changes. Such mechanisms result in high overhead page faults to access remote data and high dirty data amplification when tracking changes to cached data at page-granularity. In this paper, we propose a fundamentally new approach for disaggregated memory systems, based on the observation that we can use local cache coherence to track applications' memory accesses transparently, without code changes, at cache-line granularity. This simple idea (1) eliminates page faults from the application critical path when accessing remote data, and (2) decouples the application memory access tracking from the virtual memory page size, enabling cache-line granularity dirty data tracking and eviction. Using this observation, we implemented a new software runtime for disaggregated memory that improves average memory access time and reduces dirty data amplification1.

Simulation Study of Influencing Factors of Immersion Phase-Change Cooling Technology for Data Center Servers

The immersion phase-change cooling technology utilizes the latent heat of the cooling liquid to dissipate heat by directly contacting the cooling liquid with the heat-generating electronic chip, which can meet the cooling requirements of current high heat flux density data centers. In this paper, the effect of different factors on the heat dissipation performance of immersion phase-change cooling technology was explored through numerical simulation. The results show that, under certain power conditions, the inlet temperature and flow rate of the cooling water in the condensation module, as well as the different arrangement of servers, have a significant impact on the heat dissipation performance of the entire system. The inlet water temperature mainly affects the chip temperature after stabilization. With the decrease in the inlet temperature, the chip surface temperature decreases significantly. The inlet water flow rate mainly affects the time required for the heat exchange to reach the desired temperature. With the increase in the inlet flow rate, the required cooling time is shortened. As the spacing between servers increases, the thermal safety and stability of the entire system increase. When the spacing between servers increases from 5 mm to 15 mm, the highest temperature and the temperature uniformity coefficient between the systems decrease significantly. When the spacing increases from 15 mm to 25 mm, the highest temperature and the temperature uniformity coefficient decrease slightly. These results can provide useful information for the designers of immersion phase-change cooling systems to improve the cooling efficiency of data centers, save energy, and ensure the safe operation of related computers, servers, and communication systems.

Energy Consumption and Energy Saving Analysis of Air-Conditioning Systems of Data Centers in Typical Cities in China

The energy consumption status and energy saving potential of the air-conditioning systems of data centers in seven typical cities in China (Xinjiang, Beijing, Jinan, Shanghai, Nanning, Guilin, and Haikou), representing diverse climate regions, were studied. The power usage effectiveness (PUE) and cooling load factor (CLF) were taken as the evaluation indicators. First, the energy consumption situations of the existing air-conditioning systems were analyzed using an internship survey. Second, the meteorological data throughout the year for the seven cities were statistically analyzed. Then, two energy saving renovation schemes were proposed. The operating hours under different operating modes in the seven cities were calculated, and the PUE and energy saving potential of the two energy saving schemes were evaluated by taking the production of 15 °C chilled water as an example. This study provides an overall picture of the energy utilization status in the current Chinese data center market and provides solutions for improving the design of air-conditioning systems, with energy saving benefits.

Load Balancing With Deadline-Driven Parallel Data Transmission in Data Center Networks

With the explosive growth of the Internet of Things (IoT), an increasing amount of sensor data generated by soft real-time IoT applications has been moved to data centers for storage and data analysis. Large amounts of these data are required to be processed within a given deadline to ensure application performance. Therefore, meeting the transmission deadlines of data flows for soft real-time applications has always been crucial yet challenging to current data centers. Recent progress has demonstrated that adopting parallel data transmission over multipath data center network combining with effective load balancing can achieve a high bisection network bandwidth, thus speeding up the network transfer of data flows. Nevertheless, the deadline miss ratios (DMRs) of these flows are not lowered as expected since the existing load balancing schemes are naturally agnostic to the deadline requirement. They are either unable to reroute traffic flexibly or aimlessly reroute these deadline-restrained flows, regardless of their urgent levels and path conditions. To address these inefficiencies, we propose a deadline-aware load-balancing scheme, namely, DLB, which perceives the deadline requirements and helps the urgent flows to timely switch to those faster transmission paths to complete quickly. Specifically, DLB computes the urgent level for each flow in real time to judge if the switch needs to make proactive rerouting. When a flow is nonurgent, DLB does not proactively change its transmission path, leaving more available paths to those flows with higher urgent levels. When a flow becomes extremely urgent, it immediately switches to those light-loaded paths to finish its data transmission before its deadline as far as possible. Experimental results of NS2 simulations and real testbed implementations show that DLB reduces the DMRs by up to 50% compared to the state-of-the-art data center load-balancing schemes, while only induces trivial overhead during deployment.

ReSAW: a reconfigurable and picosecond-synchronized optical data center network based on an AWGR and the WR protocol

The explosive growth of data center (DC) traffic imposes unprecedented challenges on the current electrical switch-based data center networks (DCNs) with the bottleneck of limited bandwidth and high latency. Benefitting from transparency to the data rate and format, optical switching with theoretically infinite bandwidth could overcome the bandwidth bottleneck of electrical switching DCNs. However, DCNs normally deploy multitenant applications with a variety of DC traffic. It is hard to reconfigure the optical interconnections in real-time to provide adaptable bandwidth to the traffic with heterogeneous characteristics. Moreover, to improve the bandwidth utilization, statistical multiplexing is generally deployed in optical DCNs to forward the traffic flow in a time slot, which requires the network time to be precisely synchronized to all network nodes. For fast optical switches with a nanosecond switching configuration time, overall end node times must be synchronized at subnanosecond magnitude. In this paper, we propose and experimentally investigate a reconfigurable and picosecond-synchronized optical DCN (ReSAW) based on an arrayed waveguide grating router (AWGR) and the White Rabbit (WR) protocol. A scheduler based on a distributed field-programmable gate array is implemented in the proposed ReSAW to realize flexible wavelength configuration by controlling the fast laser array based on semiconductor optical amplifiers (SOAs) according to time slot and traffic priority. Moreover, the WR protocol is implemented in optical DCNs for what we believe is the first time to synchronize the time of the distributed top of racks (ToRs). The experimental demonstration validates that ReSAW achieves an average end-to-end latency of 317.44 ns and a precise synchronized time with an average 386 ps skew. When the load is 0.4, the packet loss after ReSAW reconfiguration is less than 1.83 × 10 − 6 , and the network latency is less than 1.73 µs. Based on the experimental parameters and results, the OMNeT++ simulation model is built to further verify the reconfigurability and scalability of the ReSAW network. Results show that the packet loss rate and latency performance increase 8.24% and 12.47%, respectively, at a load of 0.6 as the ReSAW network scales from 2560 to 40,960 servers compared to before the reconfiguration.

FastCredit: Expediting credit-based congestion control in datacenters

Receiver-driven transports that leverage hop-by-hop credits have the merits of fast convergence, short queueing, and strong congestion avoidability, making them high-efficiency in managing congestion in current high-speed data centers. However, these transport protocols treat long and latency-sensitive short flows equally, increasing flow completion times and affecting user experiences. Although flow scheduling mechanisms have been studied extensively, they are hard to be deployed into hop-by-hop credit-based transports. The root cause is that most traditional flow scheduling mechanisms mainly work in the long queue containing flows of various sizes, while credit-based proactive transports maintain the extremely short bounded queue.Based on this observation, this paper presents FastCredit, making the first attempt to expedite short-flow scheduling in hop-by-hop credit-based congestion control. In FastCredit, we schedule credit transmission at both receivers and switches to indirectly perform data flow scheduling, and we develop an ECN-based credit rate control mechanism to mitigate the problems caused by credit loss. Compared to the state-of-the-art credit-based transport protocol, FastCredit reduces average flow completion time to 0.71x and greatly improves the short flow transmission latency to 0.45x in realistic workloads. Especially, FastCredit reduces average flow completion time to 0.72x under incast circumstances and 0.56x in many-to-one traffic mode. Moreover, FastCredit still maintains the advantages of the short queue and high throughput and essentially reduces credit loss by employing credit rate control.

Green Energy Forecast-Based Bi-Objective Scheduling of Tasks Across Distributed Clouds

Current large-scale green cloud data centers (GCDCs) tend to consume a huge amount of energy and generate enormous carbon emissions. Existing studies have tried to solve this problem by either realizing prediction of green energy, or optimizing task scheduling. In contrast, this work seamlessly combines green energy prediction and task scheduling to jointly optimize revenue and energy cost of GCDCs. Specifically, this work designs a prediction method, named Savitzky-Golay and Long Short-Term Memory network (SG-LSTM), to realize noise filtering and forecast green energy. Based on such prediction, a bi-objective optimization method, named Decomposition-based Multi-objective evolutionary algorithm with Gaussian mutation and Crowding distance (DMGC), is developed to optimize the revenue and energy cost of GCDCs. Its performance is demonstrated over real-life datasets including Google cluster traces, wind speeds, solar irradiance and prices of electricity. Experimental results show that SG-LSTM outperforms its two peers, back propagation neural network and gated recurrent unit, in terms of root mean square errors and mean absolute errors. In addition, DMGC surpasses its such peers as NSGA-II, SPEA2, and MOEA/D in terms of revenue, energy cost and average execution time. Particularly, DMGC's revenue is 18%, 20% and 13.1% higher, energy cost is 16%, 19.8% and 15.2% lower, and average execution time is 60.02%, 38.47% and 24.17% lower than those of NSGA-II, SPEA2, and MOEA/D, respectively.

Task Failure Prediction in Cloud Data Centers Using Deep Learning

A large-scale cloud data center needs to provide high service reliability and availability with low failure occurrence probability. However, current large-scale cloud data centers still face high failure rates due to many reasons such as hardware and software failures, which often result in task and job failures. Such failures can severely reduce the reliability of cloud services and also occupy huge amount of resources to recover the service from failures. Therefore, it is important to predict task or job failures before occurrence with high accuracy to avoid unexpected wastage. Many machine learning and deep learning based methods have been proposed for the task or job failure prediction by analyzing past system message logs and identifying the relationship between the data and the failures. In order to further improve the failure prediction accuracy of the previous machine learning and deep learning based methods, in this article, we propose a failure prediction algorithm based on multi-layer Bidirectional Long Short Term Memory (Bi-LSTM) to identify task and job failures in the cloud. The goal of Bi-LSTM failure prediction algorithm is to predict whether the tasks and jobs are failed or completed. The trace-driven experiments show that our algorithm outperforms other state-of-art prediction methods with 93 percent accuracy and 87 percent for task failure and job failures respectively.

Energy-aware systems for real-time job scheduling in cloud data centers: A deep reinforcement learning approach

With the advantages such as high-performance, low-maintenance, and reliability, more and more companies are moving their computing infrastructures to the cloud. In the meantime, with the increasing number of users continuously submitting jobs to cloud, energy consumed by the current cloud data centers has become a major concern for cloud service providers, due to financial and environmental reasons. In this paper, we propose a deep reinforcement learning (DRL) approach to handle real-time jobs. Specifically, we focus on allocating incoming jobs to appropriate virtual machines (VMs) in a way that energy consumption can be optimized while high quality of service (QoS) can be achieved. We give the detailed design and implementation of our approach, and our experimental results demonstrate that the proposed method can achieve better performance in job success rate and average response time with less energy consumption than the current approaches, in the presence of different real-time cloud workloads.

CAPTON—Centralized All-Path Transmission for Over-Subscribed Datacenter Networks

This work addresses an often-overlooked challenge in designing datacenter traffic control schemes. Specifically, most existing schemes were designed for network topologies with non-blocking network core. While this enabled the development of elegant solutions, core networks in practice are more likely to be <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">over-subscribed</i> by the access layer network due to cost considerations. Consequently, existing traffic control schemes may perform sub-optimally in real-world over-subscribed datacenter networks. This work proposes a new centralized flow-based scheduling scheme called CAPTON which schedules flow according to remaining flow size and link capacities to avoid congestion in over-subscribed networks. CAPTON employs all-path transmission to exploit all available paths and redundant links between source and destination hosts to maximize resource utilization and eliminate the need for routing. Unlike switch-based approaches, CAPTON does not require additional support from network switches and thus can be readily deployed in current datacenters equipped with commodity switches.

DACON: a reconfigurable application-centric optical network for disaggregated data center infrastructures [Invited

To solve the issues of low resource utilization and performance bottleneck in current server-centric data center networks (DCNs), we propose and experimentally demonstrate a disaggregated application-centric optical network (DACON) for data center infrastructures based on hybrid optical switches. DACON achieves flexible provision and dynamic reconfiguration of hardware nodes exploiting the softwared-define networking (SDN) orchestration plane. A four-node SDN-enabled disaggregated prototype is implemented with a field-programmable-gate-array-based controller of hardware nodes and nanosecond optical switches, performing a minimal node-to-node network latency of 378.6 ns and zero packet loss. Based on the unmodified Linux kernel and two different applications (distributed computing and a Memcached database), the application runtime of the disaggregated prototype is investigated and compared with the server-centric architecture. Experimental results show that the disaggregated prototype performs better with Memcached database applications, achieving a1.46×faster runtime than the server-centric network at a memory node access ratio of 0.9. Based on the customized control plane orchestrator and dynamic resource reallocation, the node-to-node latency is reduced by 21% when CPU nodes access memory nodes. The scalability of DACON is then numerically assessed based on experimentally measured parameters. Results show that the intra-rack node-to-node latency is less than 404.8 ns with a 6144-node network and memory node access ratio of 0.9. Finally, the cost and power consumption are also studied and compared with current DCN architectures. Results indicate DACON saves 13.4% of the cost of an interconnect network compared with current disaggregated architecture and consumes up to 31.1% less power with respect to server-centric DCN architectures.

Ultrafast optical circuit switching for data centers using integrated soliton microcombs

Due to the slowdown of Moore’s law, it will become increasingly challenging to efficiently scale the network in current data centers utilizing electrical packet switches as data rates grow. Optical circuit switches (OCS) represent an appealing option to overcome this issue by eliminating the need for expensive and power-hungry transceivers and electrical switches in the core of the network. In particular, optical switches based on tunable lasers and arrayed waveguide grating routers are quite promising due to the use of a passive core, which increases fault tolerance and reduces management overhead. Such an OCS-network can offer high bandwidth, low network latency and an energy-efficient and scalable data center network. To support dynamic data center workloads efficiently, however, it is critical to switch between wavelengths at nanosecond (ns) timescales. Here we demonstrate ultrafast OCS based on a microcomb and semiconductor optical amplifiers (SOAs). Using a photonic integrated Si3N4 microcomb, sub-ns (<520 ps) switching along with the 25-Gbps non-return-to-zero (NRZ) and 50-Gbps four-level pulse amplitude modulation (PAM-4) burst mode data transmission is achieved. Further, we use a photonic integrated circuit comprising an Indium phosphide based SOA array and an arrayed waveguide grating to show sub-ns switching (<900 ps) along with 25-Gbps NRZ burst mode transmission providing a path towards a more scalable and energy-efficient wavelength-switched network for data centers in the post Moore’s Law era.

Load balancing with traffic isolation in data center networks

The topologies of current data center networks are typically multi-rooted trees (e.g. leaf–spine) with rich parallel paths between any pair of hosts. Recent progress has demonstrated that effective load balancing can fully utilize these parallel paths to speed up data transfer. Nonetheless, the existing load balancing designs are agnostic to the heterogeneous datacenter traffic, i.e., a mass of delay-sensitive short flows mix with a handful of throughput-oriented long flows, and casually reroute these flows regardless of path condition, thus resulting in frequent flow collisions. The short flows suffer from the problem of large queuing delay due to colliding with long flows, while the frequent collisions between long flows lead to low network utilization and serious throughput degradation. To address these inefficiencies, we propose an isolation-based load balancing scheme, namely ILB, which perceives flow collisions and dynamically isolates the long flows from short flows. Specifically, when a long flow collides with short flows in the same path, it immediately switches to another path unused by short flows to help the short ones in the previous path complete quickly. When short flows disappear, the long flow quickly occupies all its available paths to achieve high throughput. Moreover, ILB is only deployed on the leaf switch without modifying the end-host and spine switch. Experimental results of NS3 simulations show that ILB respectively reduces the average and tail flow completion time for short flows by up to 30% and 70%, as well as increases the throughput of long flows by about 1.68x over the state-of-the-art datacenter load balancing schemes.