Many applications are increasingly becoming I/O-bound. To improve scalability, analytical models of parallel I/O performance are often consulted to determine possible I/O optimizations. However, I/O performance modeling has predominantly focused on applications that directly issue I/O requests to a parallel file system or a local storage device. These I/O models are not directly usable by applications that access data through standardized I/O libraries, such as HDF5, FITS, and NetCDF, because a single I/O request to an object can trigger a cascade of I/O operations to different storage blocks. The I/O performance characteristics of applications that rely on these libraries is a complex function of the underlying data storage model, user-configurable parameters and object-level access patterns. As a consequence, I/O optimization is predominantly an ad-hoc process that is performed by application developers, who are often domain scientists with limited desire to delve into nuances of the storage hierarchy of modern computers.This paper presents an analytical cost model to predict the end-to-end execution time of applications that perform I/O through established array management libraries. The paper focuses on the HDF5 and Zarr array libraries, as examples of I/O libraries with radically different storage models: HDF5 stores every object in one file, while Zarr creates multiple files to store different objects. We find that accessing array objects via these I/O libraries introduces new overheads and optimizations. Specifically, in addition to I/O time, it is crucial to model the cost of transforming data to a particular storage layout (memory copy cost), as well as model the benefit of accessing a software cache. We evaluate the model on real applications that process observations (neuroscience) and simulation results (plasma physics). The evaluation on three HPC clusters reveals that I/O accounts for as little as 10% of the execution time in some cases, and hence models that only focus on I/O performance cannot accurately capture the performance of applications that use standard array storage libraries. In parallel experiments, our model correctly predicts the fastest storage library between HDF5 and Zarr 94% of the time, in contrast with 70% of the time for a cutting-edge I/O model.
more »
« less
Mapping Datasets to Object Storage System
Access libraries such as ROOT[1] and HDF5[2] allow users to interact with datasets using high level abstractions, like coordinate systems and associated slicing operations. Unfortunately, the implementations of access libraries are based on outdated assumptions about storage systems interfaces and are generally unable to fully benefit from modern fast storage devices. For example, access libraries often implement buffering and data layout that assume that large, single-threaded sequential access patterns are causing less overall latency than small parallel random access: while this is true for spinning media, it is not true for flash media. The situation is getting worse with rapidly evolving storage devices such as non-volatile memory and ever larger datasets. This project explores distributed dataset mapping infrastructures that can integrate and scale out existing access libraries using Ceph’s extensible object model, avoiding re-implementation or even modifications of these access libraries as much as possible. These programmable storage extensions coupled with our distributed dataset mapping techniques enable: 1) access library operations to be offloaded to storage system servers, 2) the independent evolution of access libraries and storage systems and 3) fully leveraging of the existing load balancing, elasticity, and failure management of distributed storage systems like Ceph. They also create more opportunities to conduct storage server-local optimizations specific to storage servers. For example, storage servers might include local key/value stores combined with chunk stores that require different optimizations than a local file system. As storage servers evolve to support new storage devices like non-volatile memory, these server-local optimizations can be implemented while minimizing disruptions to applications. We will report progress on the means by which distributed dataset mapping can be abstracted over particular access libraries, including access libraries for ROOT data, and how we address some of the challenges revolving around data partitioning and composability of access operations.
more »
« less
- NSF-PAR ID:
- 10275751
- Date Published:
- Journal Name:
- EPJ web of conferences
- Volume:
- 245
- Issue:
- 04037
- ISSN:
- 2100-014X
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
With the ever-increasing dataset sizes, several file formats such as Parquet, ORC, and Avro have been developed to store data efficiently, save the network, and interconnect bandwidth at the price of additional CPU utilization. However, with the advent of networks supporting 25-100 Gb/s and storage devices delivering 1,000,000 reqs/sec, the CPU has become the bottleneck trying to keep up feeding data in and out of these fast devices. The result is that data access libraries executed on single clients are often CPU-bound and cannot utilize the scale-out benefits of distributed storage systems. One attractive solution to this problem is to offload data-reducing processing and filtering tasks to the storage layer. However, modifying legacy storage systems to support compute offloading is often tedious and requires an extensive understanding of the system internals. Previous approaches re-implemented functionality of data processing frameworks and access libraries for a particular storage system, a duplication of effort that might have to be repeated for different storage systems. This paper introduces a new design paradigm that allows extending programmable object storage systems to embed existing, widely used data processing frameworks and access libraries into the storage layer with no modifications. In this approach, data processing frameworks and access libraries can evolve independently from storage systems while leveraging distributed storage systems’ scale-out and availability properties. We present Skyhook, an example implementation of our design paradigm using Ceph, Apache Arrow, and Parquet. We provide a brief performance evaluation of Skyhook and discuss key results.more » « less
-
The Skyhook Data Management project (SkyhookDM.com) at the Center for Research in Open Source Software (cross.ucsc.edu) at UC Santa Cruz implements customized extensions through Ceph's object class interface that enables offloading database operations to the storage system. In our previous Vault '19 talk, we showed how SkyhookDM can transparently scale out databases. The SkyhookDM Ceph extensions are an example of our 'programmable storage' research efforts at UCSC, and can be accessed through commonly available external/foreign table database interfaces. Utilizing fast in-memory serialization libraries such as Google Flatbuffers and Apache Arrow, SkyhookDM currently implements common database functions such as SELECT, PROJECT, AGGREGATE, and indexing inside Ceph, along with lower-level data manipulations such as transforming data from row to column formats on RADOS servers. In this talk, we will present three of our latest developments on the SkyhookDM project since Vault '19. First, SkyhookDM can be used to also offload operations of access libraries that support plugins for backends, such as HDF5 and its Virtual Object Layer. Second, in addition to row-oriented data format using Google's Flatbuffers, we have added support for column-oriented data formats using the Apache Arrow library within our Ceph extensions. Third, we added dynamic switching between row and column data formats within Ceph objects, a first step towards physical design management in storage systems, similar to physical design tuning in database systems.more » « less
-
Despite the recent success of Graph Neural Networks (GNNs), training GNNs on large graphs remains challenging. The limited resource capacities of the existing servers, the dependency between nodes in a graph, and the privacy concern due to the centralized storage and model learning have spurred the need to design an effective distributed algorithm for GNN training. However, existing distributed GNN training methods impose either excessive communication costs or large memory overheads that hinders their scalability. To overcome these issues, we propose a communication-efficient distributed GNN training technique named (LLCG). To reduce the communication and memory overhead, each local machine in LLCG first trains a GNN on its local data by ignoring the dependency between nodes among different machines, then sends the locally trained model to the server for periodic model averaging. However, ignoring node dependency could result in significant performance degradation. To solve the performance degradation, we propose to apply on the server to refine the locally learned models. We rigorously analyze the convergence of distributed methods with periodic model averaging for training GNNs and show that naively applying periodic model averaging but ignoring the dependency between nodes will suffer from an irreducible residual error. However, this residual error can be eliminated by utilizing the proposed global corrections to entail fast convergence rate. Extensive experiments on real-world datasets show that LLCG can significantly improve the efficiency without hurting the performance. One-sentence Summary: We propose LLCG a communication efficient distributed algorithm for training GNNs.more » « less
-
Replication is essential for fault-tolerance. However, in in-memory systems, it is a source of high overhead. Remote direct memory access (RDMA) is attractive to create redundant copies of data, since it is low-latency and has no CPU overhead at the target. However, existing approaches still result in redundant data copying and active receivers. To ensure atomic data transfers, receivers check and apply only fully received messages. Tailwind is a zero-copy recovery-log replication protocol for scale-out in-memory databases. Tailwind is the first replication protocol that eliminates all CPU-driven data copying and fully bypasses target server CPUs, thus leaving backups idle. Tailwind ensures all writes are atomic by leveraging a protocol that detects incomplete RDMA transfers. Tailwind substantially improves replication throughput and response latency compared with conventional RPC-based replication. In symmetric systems where servers both serve requests and act as replicas, Tailwind also improves normal-case throughput by freeing server CPU resources for request processing. We implemented and evaluated Tailwind on RAMCloud, a low-latency in-memory storage system. Experiments show Tailwind improves RAMCloud's normal-case request processing throughput by 1.7x. It also cuts down writes median and 99th percentile latencies by 2x and 3x respectively.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- https://doi.org/202024504037
