Interactive notebook programming is universal in modern ML and AI workflows, with interactive deep learning training (IDLT) emerging as a dominant use case.To ensure responsiveness, platforms like Jupyter and Colab reserve GPUs for long-running notebook sessions, despite their intermittent and sporadic GPU usage, leading to extremely low GPU utilization and prohibitively high costs.In this paper, we introduce NotebookOS, a GPU-efficient notebook platform tailored for the unique requirements of IDLT.NotebookOS employs replicated notebook kernels with Raft-synchronized replicas distributed across GPU servers.To optimize GPU utilization, NotebookOS oversubscribes server resources, leveraging high inter-arrival times in IDLT workloads, and allocates GPUs only during active cell execution.It also supports replica migration and automatic cluster scaling under high load.Altogether, this design enables interactive training with minimal delay.In evaluation on production workloads, NotebookOS saved over 1,187 GPU hours in 17.5 hours of real-world IDLT, while significantly improving interactivity.

Cold start delays are a main pain point for today's FaaS (Function-as-a-Service) platforms. A widely used mitigation strategy is keeping recently invoked function containers alive in memory to enable warm starts with minimal overhead. This paper identifies new challenges that state-of-the-art FaaS keep-alive policies neglect. These challenges are caused by concurrent function invocations, a common FaaS workload behavior. First, concurrent requests present a tradeoff between reusing busy containers (delayed warm starts) versus cold-starting containers. Second, concurrent requests cause imbalanced evictions of containers that will be reused shortly thereafter. To tackle the challenges, we propose a novel serverless function container orchestration algorithm called CIDRE. CIDRE makes informed decisions to speculatively choose between a delayed warm start and a cold start under concurrency-driven function scaling. CIDRE uses both fine-grained container-level and coarse-grained concurrency information to make balanced eviction decisions. We evaluate CIDRE extensively using two production FaaS workloads. Results show that CIDRE reduces the cold start ratio and the average invocation overhead by up to 75.1% and 39.3% compared to state-of-the-art function keep-alive policies.

Web 3.0 is redefining the current Web (Web 2.0) with a focus on data and governance decentralization. The InterPlanetary File System (IPFS) exemplifies this shift. However, it faces a trade-off between decentralization and performance: prior studies have shown IPFS's performance degradations but fail to diagnose root causes or deliver actionable fixes.

Despite the recent success of Graph Neural Networks (GNNs), it remains challenging to train GNNs on large-scale graphs due to neighbor explosions. As a remedy, distributed computing becomes a promising solution by leveraging abundant computing resources (e.g., GPU). However, the node dependency of graph data increases the difficulty of achieving high concurrency in distributed GNN training, which suffers from the massive communication overhead. To address it, Historical value approximation is deemed a promising class of distributed training techniques. It utilizes an offline memory to cache historical information (e.g., node embedding) as an affordable approximation of the exact value and achieves high concurrency. However, such benefits come at the cost of involving dated training information, leading to staleness, imprecision, and convergence issues. To overcome these challenges, this paper proposes SAT (Staleness-Alleviated Training), a novel and scalable distributed GNN training framework that reduces the embedding staleness adaptively. The key idea of SAT is to model the GNN’s embedding evolution as a temporal graph and build a model upon it to predict future embedding, which effectively alleviates the staleness of the cached historical embedding. We propose an online algorithm to train the embedding predictor and the distributed GNN alternatively and further provide a convergence analysis. Empirically, we demonstrate that SAT can effectively reduce embedding staleness and thus achieve better performance and convergence speed on multiple large-scale graph datasets. Code available at https://github.com/Ziyang-Yu/SAT.

The InterPlanetary File System (IPFS) is a pioneering effort for Web 3.0, well-known for its decentralized infrastructure. However, some recent studies have shown that IPFS exhibits a high degree of centralization and has integrated centralized components for improved performance. While this change contradicts the core decentralized ethos of IPFS and introduces risks of hurting the data replication level and thus availability, it also opens some opportunities for better data management and cost savings through deduplication.

Federated learning (FL) has emerged as a new paradigm of machine learning (ML) with the goal of collaborative learning on the vast pool of private data available across distributed edge devices. The focus of most existing works in FL systems has been on addressing the challenges of computation and communication heterogeneity inherent in training with edge devices. However, the crucial impact of I/O and the role of limited on-device storage has not been explored fully in FL context. Without policies to exploit the on-device storage for placement of client data samples, and schedule clients based on I/O benefits, FL training can lead to inefficiencies, such as increased training time and impacted accuracy convergence.

As the number of pre-trained machine learning (ML) models is growing exponentially, data reduction tools are not catching up. Existing data reduction techniques are not specifically designed for pre-trained model (PTM) dataset files. This is largely due to a lack of understanding of the patterns and characteristics of these datasets, especially those relevant to data reduction and compressibility. This paper presents the first, exhaustive analysis to date of PTM datasets on storage compressibility. Our analysis spans different types of data reduction and compression techniques, from hash-based data deduplication, data similarity detection, to dictionary-coding compression. Our analysis explores these techniques at three data granularity levels, from model layers, model chunks, to model parameters. We draw new observations that indicate that modern data reduction tools are not effective when handling PTM datasets. There is a pressing need for new compression methods that take into account PTMs' data characteristics for effective storage reduction. Motivated by our findings, we design Elf, a simple yet effective, error-bounded, lossy floating-point compression method. Elf transforms floating-point parameters in such a way that the common exponent field of the transformed parameters can be completely eliminated to save storage space. We develop Elves, a compression framework that integrates Elf along with several other data reduction methods. Elves uses the most effective method to compress PTMs that exhibit different patterns. Evaluation shows that Elves achieves an overall compression ratio of 1.52×, which is 1.31×, 1.32× and 1.29× higher than a general-purpose compressor (zstd), an error-bounded lossy compressor (SZ3), and the uniform model quantization, respectively, with negligible model accuracy loss.

While graph neural networks (GNNs) are popular in the deep learning community, they suffer from several challenges including over-smoothing, over-squashing, and gradient vanishing. Recently, a series of models have attempted to relieve these issues by first augmenting the node features and then imposing node-wise functions based on multilayer perceptron (MLP), which are widely referred to as graph-augmented MLP (GA-MLP) models. However, while GA-MLP models enjoy deeper architectures for better accuracy, their efficiency largely deteriorates. Moreover, popular acceleration techniques such as stochastic-version or data-parallelism cannot be effectively applied due to the dependency among samples (i.e., nodes) in graphs. To address these issues, in this article, instead of data parallelism, we propose a parallel graph deep learning Alternating Direction Method of Multipliers (pdADMM-G) framework to achieve model parallelism: parameters in each layer of GA-MLP models can be updated in parallel. The extended pdADMM-G-Q algorithm reduces communication costs by introducing the quantization technique. Theoretical convergence to a (quantized) stationary point of the pdADMM-G algorithm and the pdADMM-G-Q algorithm is provided with a sublinear convergence rate o(1/k) , where k is the number of iterations. Extensive experiments demonstrate the convergence of two proposed algorithms. Moreover, they lead to a more massive speedup and better performance than all state-of-the-art comparison methods on nine benchmark datasets. Last but not least, the proposed pdADMM-G-Q algorithm reduces communication overheads by up to 45% without loss of performance. Our code is available at https://github.com/xianggebenben/pdADMM-G.

The metadata service (MDS) sits on the critical path for distributed file system (DFS) operations, and therefore it is key to the overall performance of a large-scale DFS. Common "serverful" MDS architectures, such as a single server or cluster of servers, have a significant shortcoming: either they are not scalable, or they make it difficult to achieve an optimal balance of performance, resource utilization, and cost. A modern MDS requires a novel architecture that addresses this shortcoming.

Federated Learning (FL) is a machine learning approach that addresses privacy and data transfer costs by computing data at the source. It’s particularly popular for Edge and IoT applications where the aggregator server of FL is in resource-capped edge data centers for reducing communication costs. Existing cloud-based aggregator solutions are resource-inefficient and expensive at the Edge, leading to low scalability and high latency. To address these challenges, this study compares prior and new aggregation methodologies under the changing demands of IoT and Edge applications. This work is the first to propose an adaptive FL aggregator at the Edge, enabling users to manage the cost and efficiency trade-off. An extensive comparative analysis demonstrates that the design improves scalability by up to 4$\times$, time efficiency by 8$\times$, and reduces costs by more than 2$\times$ compared to extant cloud-based static methodologies.

Learned Index Structures (LIS) view a sorted index as a model that learns the data distribution, takes a data element key as input, and outputs the predicted position of the key. The original LIS can only handle lookup operations with no support for updates, rendering it impractical to use for typical workloads. To address this limitation, recent studies have focused on designing efficient dynamic learned indexes. ALEX, as the first and one of the representative dynamic learned index structures, enables dynamism by incorporating a series of design choices, including adaptive key space partitioning, dynamic model retraining, and sophisticated engineering and policies that prioritize read/write performance. While these design choices offer improved average-case performance, the emphasis on flexibility and performance increases the attack surface by allowing adversarial behaviors that maximize ALEX's memory space and time complexity in worst-case scenarios. In this work, we present the first systematic investigation of algorithmic complexity attacks (ACAs) targeting the worst-case scenarios of ALEX. We introduce new ACAs that fall into two categories, space ACAs and time ACAs, which target the memory space and time complexity, respectively. First, our space ACA on data nodes exploits ALEX's gapped array layout and uses Multiple-Choice Knapsack (MCK) to generate an optimal adversarial insertion plan for maximizing the memory consumption at the data node level. Second, our space ACA on internal nodes exploits ALEX's catastrophic cost mitigation mechanism, causing an out-of-memory (OOM) error with only a few hundred adversarial insertions. Third, our time ACA generates pathological insertions to increase the disparity between the actual key distribution and the linear models of data nodes, deteriorating the runtime performance by up to 1, 641× compared to ALEX operating under legitimate workloads.

Cloud object storage such as AWS S3 is cost-effective and highly elastic but relatively slow, while high-performance cloud storage such as AWS ElastiCache is expensive and provides limited elasticity. We present a new cloud storage service called ServerlessMemory, which stores data using the memory of serverless functions. ServerlessMemory employs a sliding-window-based memory management strategy inspired by the garbage collection mechanisms used in the programming language to effectively segregate hot/cold data and provides fine-grained elasticity, good performance, and a pay-per-access cost model with extremely low cost. We then design and implement InfiniStore, a persistent and elastic cloud storage system, which seamlessly couples the function-based ServerlessMemory layer with a persistent, inexpensive cloud object store layer. InfiniStore enables durability despite function failures using a fast parallel recovery scheme built on the auto-scaling functionality of a FaaS (Function-as-a-Service) platform. We evaluate InfiniStore extensively using both microbenchmarking and two real-world applications. Results show that InfiniStore has more performance benefits for objects larger than 10 MB compared to AWS ElastiCache and Anna, and InfiniStore achieves 26.25% and 97.24% tenant-side cost reduction compared to InfiniCache and ElastiCache, respectively.

Serverless computing enables a new way of building and scaling cloud applications by allowing developers to write fine-grained serverless or cloud functions. The execution duration of a cloud function is typically short-ranging from a few milliseconds to hundreds of seconds. However, due to resource contentions caused by public clouds' deep consolidation, the function execution duration may get significantly prolonged and fail to accurately account for the function's true resource usage. We observe that the function duration can be highly unpredictable with huge amplification of more than 50× for an open-source FaaS platform (OpenLambda). Our experiments show that the OS scheduling policy of cloud functions' host server can have a crucial impact on performance. The default Linux scheduler, CFS (Completely Fair Scheduler), being oblivious to workloads, frequently context-switches short functions, causing a turnaround time that is much longer than their service time. We propose SFS (Smart Function Scheduler), which works entirely in the user space and carefully orchestrates existing Linux FIFO and CFS schedulers to approximate Shortest Remaining Time First (SRTF). SFS uses two-level scheduling that seamlessly combines a new FILTER policy with Linux CFS, to trade off increased duration of long functions for significant performance improvement for short functions. We implement SFS in the Linux user space and port it to OpenLambda. Evaluation results show that SFS significantly improves short functions' duration with a small impact on relatively longer functions, compared to CFS.

Federated learning (FL) involves multiple distributed devices jointly training a shared model without any of the participants having to reveal their local data to a centralized server. Most of previous FL approaches assume that data on devices are fixed and stationary during the training process. However, this assumption is unrealistic because these devices usually have varying sampling rates and different system configurations. In addition, the underlying distribution of the device data can change dynamically over time, which is known as concept drift. Concept drift makes the learning process complicated because of the inconsistency between existing and upcoming data. Traditional concept drift handling techniques such as chunk based and ensemble learning-based methods are not suitable in the federated learning frameworks due to the heterogeneity of local devices. We propose a novel approach, FedConD, to detect and deal with the concept drift on local devices and minimize the effect on the performance of models in asynchronous FL. The drift detection strategy is based on an adaptive mechanism which uses the historical performance of the local models. The drift adaptation is realized by adjusting the regularization parameter of objective function on each local device. Additionally, we design a communication strategy on the server side to select local updates in a prudent fashion and speed up model convergence. Experimental evaluations on three evolving data streams and two image datasets show that FedConD detects and handles concept drift, and also reduces the overall communication cost compared to other baseline methods.

Containerization is becoming increasingly popular, but unfortunately, containers often fail to deliver the anticipated performance with the allocated resources. In this paper, we first demonstrate the performance variance and degradation are significant (by up to 5x) in a multi-tenant environment where containers are co-located. We then investigate the root cause of such performance degradation. Contrary to the common belief that such degradation is caused by resource contention and interference, we find that there is a gap between the amount of CPU a container reserves and actually gets. The root cause lies in the design choices of today's Linux scheduling mechanism, which we call Forced Runqueue Sharing and Phantom CPU Time. In fact, there are fundamental conflicts between the need to reserve CPU resources and Completely Fair Scheduler's work-conserving nature, and this contradiction prevents a container from fully utilizing its requested CPU resources. As a proof-of-concept, we implement a new resource configuration mechanism atop the widely used Kubernetes and Linux to demonstrate its potential benefits and shed light on future scheduler redesign. Our proof-of-concept, compared to the existing scheduler, improves the performance of both batch and interactive containerized apps by up to 5.6x and 13.7x.

Federated learning (FL) involves training a model over massive distributed devices, while keeping the training data localized and private. This form of collaborative learning exposes new tradeoffs among model convergence speed, model accuracy, balance across clients, and communication cost, with new challenges including: (1) straggler problem---where clients lag due to data or (computing and network) resource heterogeneity, and (2) communication bottleneck---where a large number of clients communicate their local updates to a central server and bottleneck the server. Many existing FL methods focus on optimizing along only one single dimension of the tradeoff space. Existing solutions use asynchronous model updating or tiering-based, synchronous mechanisms to tackle the straggler problem. However, asynchronous methods can easily create a communication bottleneck, while tiering may introduce biases that favor faster tiers with shorter response latencies.

Executing complex, burst-parallel, directed acyclic graph (DAG) jobs poses a major challenge for serverless execution frameworks, which will need to rapidly scale and schedule tasks at high throughput, while minimizing data movement across tasks. We demonstrate that, for serverless parallel computations, decentralized scheduling enables scheduling to be distributed across Lambda executors that can schedule tasks in parallel, and brings multiple benefits, including enhanced data locality, reduced network I/Os, automatic resource elasticity, and improved cost effectiveness. We describe the implementation and deployment of our new serverless parallel framework, called Wukong, on AWS Lambda. We show that Wukong achieves near-ideal scalability, executes parallel computation jobs up to 68.17X faster, reduces network I/O by multiple orders of magnitude, and achieves 92.96% tenant-side cost savings compared to numpywren.

Federated Learning (FL) enables learning a shared model acrossmany clients without violating the privacy requirements. One of the key attributes in FL is the heterogeneity that exists in both resource and data due to the differences in computation and communication capacity, as well as the quantity and content of data among different clients. We conduct a case study to show that heterogeneity in resource and data has a significant impact on training time and model accuracy in conventional FL systems. To this end, we propose TiFL, a Tier-based Federated Learning System, which divides clients into tiers based on their training performance and selects clients from the same tier in each training round to mitigate the straggler problem caused by heterogeneity in resource anddata quantity. To further tame the heterogeneity caused by non-IID (Independent and Identical Distribution) data and resources, TiFL employs an adaptive tier selection approach to update the tiering on-the-fly based on the observed training performance and accuracy. We prototype TiFL in a FL testbed following Google's FL architecture and evaluate it using the state-of-the-art FL benchmarks. Experimental evaluation shows that TiFL outperforms the conventional FL in various heterogeneous conditions. With the proposed adaptive tier selection policy, we demonstrate that TiFL achieves much faster training performance while achieving the same or better test accuracy across the board.

Alternating Direction Method of Multipliers (ADMM) has recently been proposed as a potential alternative optimizer to the Stochastic Gradient Descent(SGD) for deep learning problems. This is because ADMM can solve gradient vanishing and poor conditioning problems. Moreover, it has shown good scalability in many large-scale deep learning applications. However, there still lacks a parallel ADMM computational framework for deep neural networks because of layer dependency among variables. In this paper, we propose a novel parallel deep learning ADMM framework (pdADMM) to achieve layer parallelism: parameters in each layer of neural networks can be updated independently in parallel. The convergence of the proposed pdADMM to a critical point is theoretically proven under mild conditions. The convergence rate of the pdADMM is proven to be o(1/k) where k is the number of iterations. Extensive experiments on six benchmark datasets demonstrated that our proposed pdADMM can lead to more than 10 times speedup for training large-scale deep neural networks, and outperformed most of the comparison methods. Our code is available at: https://github.com/xianggebenben/pdADMM.

Enterprise KV stores are often not well suited for HPC applications, and thus cumbersome end-to-end KV design customization is required to meet the needs of modern HPC applications. To this end, in this article we present bespoKV, an adaptive, extensible, and scale-out KV store framework. bespoKV decouples the KV store design into the control plane for distributed management and the data plane for local data store. For the control plane, bespoKVprovides pre-built modules, called controlets, supporting common distributed functionalities (e.g., replication, consistency, and topology) and their various combinations. This decoupling allows bespoKV to take a user-provided single-server KV store, called a datalet, and transparently enables a scalable and fault-tolerant distributed KV store service. The resulting distributed stores are also adaptive to consistency or topology requirement changes and can be easily extended for new types of services. Such specializations enable innovative uses of KV stores in HPC applications, especially for emerging applications that utilize KV-friendly workloads. We evaluate bespoKV in a local testbed as well as in a public cloud settings. Experiments show that bespoKV-enabled distributed KV stores scale horizontally to a large number of nodes, and performs comparably and sometimes 1.2× to 2.6× better than the state-of-the-art systems.

Docker container images are typically stored in a centralized registry to allow easy sharing of images. However, with the growing popularity of containerized software, the number of images that a registry needs to store and the rate of requests it needs to serve are increasing rapidly. Current registry design requires hosting registry services across multiple loosely connected servers with different roles such as load balancers, proxies, registry servers, and object storage servers. Due to the various individual components, registries are hard to scale and benefits from optimizations such as caching are limited. In this paper we propose, implement, and evaluate BOLT-a new hyperconverged design for container registries. In BOLT, all registry servers are part of a tightly connected cluster and play the same consolidated role: each registry server caches images in its memory, stores images in its local storage, and provides computational resources to process client requests. The design employs a custom consistent hashing function to take advantage of the layered structure and addressing of images and to load balance requests across different servers. Our evaluation using real production workloads shows that BOLT outperforms the conventional registry design significantly and improves latency by an order of magnitude and throughput by up to 5x. Compared to state-of-the-art, BOLT can utilize cache space more efficiently and serve up to 35% more requests from its cache. Furthermore, BOLT scales linearly and recovers from failure recovery without significant performance degradation.

With our increasing reliance on cloud computing, accurate resource allocation of virtual machines (or domains) in the cloud have become more and more important. However, the current design of hypervisors (or virtual machine monitors) fails to accurately allocate resources to the domains in the virtualized environment. In this paper, we claim the root cause is that the protection scope is erroneously used as the resource scope for a domain in the current virtualization design. Such design flaw prevents the hypervisor from accurately accounting resource consumption of each domain. In this paper, using virtual CPUs as a container we propose to redefine the resource scope of a domain, so that the new resource scope is aligned with all the CPU consumption incurred by this domain. As a demonstration, we implement a novel system, called VASE (vCPU as a container), on top of the Xen hypervisor. Evaluations on our testbed have shown our proposed approach is effective in accounting system-wide CPU consumption incurred by domains, while introducing negligible overhead to the system.

Exponential increases in complexity and scale make variability a growing threat to sustaining HPC performance at exascale. Performance variability in HPC I/O is common, acute, and formidable. We take the first step towards comprehensively studying linear and nonlinear approaches to modeling HPC I/O system variability in an effort to demonstrate that variability is often a predictable artifact of system design. Using over 8 months of data collection on 6 identical systems, we propose and validate a modeling and analysis approach (MOANA) that predicts HPC I/O variability for thousands of software and hardware configurations on highly parallel shared-memory systems. Our findings indicate nonlinear approaches to I/O variability prediction are an order of magnitude more accurate than linear regression techniques. We demonstrate the use of MOANA to accurately predict the confidence intervals of unmeasured I/O system configurations for a given number of repeat runs - enabling users to quantitatively balance experiment duration with statistical confidence.

Warehouse-scale cloud datacenters co-locate workloads with different and often complementary characteristics for improved resource utilization. To better understand the challenges in managing such intricate, heterogeneous workloads while providing quality-assured resource orchestration and user experience, we analyze Alibaba's co-located workload trace, the first publicly available dataset with precise information about the category of each job. Two types of workload-long-running, user-facing, containerized production jobs, and transient, highly dynamic, non-containerized, and non-production batch jobs- are running on a shared cluster of 1313 machines. Through workload characterization, we find evidences that imply that one workload scheduler makes seemingly independent scheduling decisions regardless of the co-existence of the other. This upsurges an imminent need for a more integrated, global coordinating system that transparently connect multiple resource schedulers together and cohesively coordinates the multiple heterogeneous workloads for greater efficiency. Our multifaceted analysis reveals insights that we believe are useful for system designers and IT practitioners working on cluster management systems.

NAND flash-based Solid State Devices (SSDs) offer the desirable features of high performance, energy efficiency, and fast growing capacity. Thus, the use of SSDs is increasing in distributed storage systems. A key obstacle in this context is that the natural unbalance in distributed I/O workloads can result in wear imbalance across the SSDs in a distributed setting. This, in turn can have significant impact on the reliability, performance, and lifetime of the storage deployment. Extant load balancers for storage systems do not consider SSD wear imbalance when placing data, as the main design goal of such balancers is to extract higher performance. Consequently, data migration is the only common technique for tackling wear imbalance, where existing data is moved from highly loaded servers to the least loaded ones. In this paper, we explore an innovative holistic approach, Chameleon, that employs data redundancy techniques such as replication and erasure-coding, coupled with endurance-aware write offloading, to mitigate wear level imbalance in distributed SSD-based storage. Chameleon aims to balance the wear among different flash servers while meeting desirable objectives of: extending life of flash servers; improving I/O performance; and avoiding bottlenecks. Evaluation with a 50 node SSD cluster shows that Chameleon reduces the wear distribution deviation by 81% while improving the write performance by up to 33%.

The use of cloud object stores has been growing rapidly in recent years as they combine key advantages such as HTTP-based RESTful APIs, high availability, elasticity with a "pay-as-you-go" pricing model that allows applications to scale as needed. The current practice is to either use a single set of configuration parameters or rely on statically configured storage policies for a cloud object store deployment, even when the store is used to support different types of applications with evolving requirements. This crucial mismatch between the different applications requirements and capabilities of the object store is problematic and should be addressed to achieve high efficiency and performance.

The extreme latency and throughput requirements of modern web applications are driving the use of distributed in-memory object caches such as Memcached. While extant caching systems scale-out seamlessly, their use in the cloud --- with its unique cost and multi-tenancy dynamics --- presents unique opportunities and design challenges.

Enterprises are increasingly moving their big data analytics to the cloud with the goal of reducing costs without sacrificing application performance. Cloud service providers offer their tenants a myriad of storage options, which while flexible, makes the choice of storage deployment non trivial. Crafting deployment scenarios to leverage these choices in a cost-effective manner - under the unique pricing models and multi-tenancy dynamics of the cloud environment - presents unique challenges in designing cloud-based data analytics frameworks.

In-memory object caches are extensively used in today's web installations [1, 6]. Most existing systems adopt monolithic storage models and engineer optimizations on specific workload characteristics [3, 6] or operations [4, 5]. Such optimizations are insufficient as large-scale cloud workloads typically exhibit both temporal and spatial shifts - requirements vary within the same deployment over time and different parts of the same workload demonstrate different access patterns. To this end, we propose a caching tier that supports differentiated services in multiple dimensions. Since there is no best "one-size-fits-all" solution for all workload requirements, we argue that a fine-grained modular design will provide both high performance and flexibility in supporting multiple services.