اشترك بالحزمة الذهبية واحصل على وصول غير محدود شمرا أكاديميا
تسجيل مستخدم جديدJust-in-Time Dynamic-Batching
80
0
0.0
(
0
)
اسأل ChatGPT حول البحث
ﻻ يوجد ملخص باللغة العربية
Batching is an essential technique to improve computation efficiency in deep learning frameworks. While batch processing for models with static feed-forward computation graphs is straightforward to implement, batching for dynamic computation graphs such as syntax trees or social network graphs is challenging due to variable computation graph structure across samples. Through simulation and analysis of a Tree-LSTM model, we show the key trade-off between graph analysis time and batching effectiveness in dynamic batching. Based on this finding, we propose a dynamic batching method as an extension to MXNet Gluons just-in-time compilation (JIT) framework. We show empirically that our method yields up to 6.25 times speed-up on a common dynamic workload, a tree-LSTM model for the semantic relatedness task.
قيم البحث
اقرأ أيضاً
By the CAP Theorem, a distributed data storage system can ensure either Consistency under Partition (CP) or Availability under Partition (AP), but not both. This has led to a split between CP databases, in which updates are synchronous, and AP databa
ses, where they are asynchronous. However, there is no inherent reason to treat all updates identically: simply, the system should be as available as possible, and synchronised just enough for the application to be correct. We offer a principled Just-Right Consistency approach to designing such applications, reconciling correctness with availability and performance, based on the following insights:(i) The Conflict-Free Replicated Data Type (CRDTs) data model supports asynchronous updates in an intuitive and principled way.(ii) Invariants involving joint or mutually-ordered updates are compatible with AP and can be guaranteed by Transactional Causal Consistency, the strongest consistency model that does not compromise availability. Regarding the remaining, CAP-sensitive invariants:(iii) For the common pattern of Bounded Counters, we provide encapsulated data type that is proven correct and is efficient; (iv) in the general case, static analysis can identify when synchronisation is not necessary for correctness.Our Antidote cloud database system supports CRDTs, Transactional Causal Consistency and the Bounded Counter data type. Support tools help design applications by static analysis and proof of CAP-sensitive invariants. This system supports industrial-grade applications and has been tested experimentally with hundreds of servers across several geo-distributed data centres.
The dynamic scaling of distributed computations plays an important role in the utilization of elastic computational resources, such as the cloud. It enables the provisioning and de-provisioning of resources to match dynamic resource availability and
demands. In the case of distributed graph processing, changing the number of the graph partitions while maintaining high partitioning quality imposes serious computational overheads as typically a time-consuming graph partitioning algorithm needs to execute each time repartitioning is required. In this paper, we propose a dynamic scaling method that can efficiently change the number of graph partitions while keeping its quality high. Our idea is based on two techniques: preprocessing and very fast edge partitioning, called graph edge ordering and chunk-based edge partitioning, respectively. The former converts the graph data into an ordered edge list in such a way that edges with high locality are closer to each other. The latter immediately divides the ordered edge list into an arbitrary number of high-quality partitions. The evaluation with the real-world billion-scale graphs demonstrates that our proposed approach significantly reduces the repartitioning time, while the partitioning quality it achieves is on par with that of the best existing static method.
Distributed Complex Event Processing (DCEP) is a commonly used paradigm to detect and act on situational changes of many applications, including the Internet of Things (IoT). DCEP achieves this using a simple specification of analytical tasks on data
streams called operators and their distributed execution on a set of infrastructure. The adaptivity of DCEP to the dynamics of IoT applications is essential and very challenging in the face of changing demands concerning Quality of Service. In our previous work, we addressed this issue by enabling transitions, which allow for the adaptive use of multiple operator placement mechanisms. In this article, we extend the transition methodology by optimizing the costs of transition and analyzing the behaviour using multiple operator placement mechanisms. Furthermore, we provide an extensive evaluation on the costs of transition imposed by operator migrations and learning, as it can inflict overhead on the performance if operated uncoordinatedly.
Many IoT systems are data intensive and are for the purpose of monitoring for fault detection and diagnosis of critical systems. A large volume of data steadily come out of a large number of sensors in the monitoring system. Thus, we need to consider
how to store and manage these data. Existing time series databases (TSDBs) can be used for monitoring data storage, but they do not have good models for describing the data streams stored in the database. In this paper, we develop a semantic model for the specification of the monitoring data streams (time series data) in terms of which sensor generated the data stream, which metric of which entity the sensor is monitoring, what is the relation of the entity to other entities in the system, which measurement unit is used for the data stream, etc. We have also developed a tool suite, SE-TSDB, that can run on top of existing TSDBs to help establish semantic specifications for data streams and enable semantic-based data retrievals. With our semantic model for monitoring data and our SE-TSDB tool suite, users can retrieve non-existing data streams that can be automatically derived from the semantics. Users can also retrieve data streams without knowing where they are. Semantic based retrieval is especially important in a large-scale integrated IoT-Edge-Cloud system, because of its sheer quantity of data, its huge number of computing and IoT devices that may store the data, and the dynamics in data migration and evolution. With better data semantics, data streams can be more effectively tracked and flexibly retrieved to help with timely data analysis and control decision making anywhere and anytime.
Ubers business is highly real-time in nature. PBs of data is continuously being collected from the end users such as Uber drivers, riders, restaurants, eaters and so on everyday. There is a lot of valuable information to be processed and many decisio
ns must be made in seconds for a variety of use cases such as customer incentives, fraud detection, machine learning model prediction. In addition, there is an increasing need to expose this ability to different user categories, including engineers, data scientists, executives and operations personnel which adds to the complexity. In this paper, we present the overall architecture of the real-time data infrastructure and identify three scaling challenges that we need to continuously address for each component in the architecture. At Uber, we heavily rely on open source technologies for the key areas of the infrastructure. On top of those open-source software, we add significant improvements and customizations to make the open-source solutions fit in Ubers environment and bridge the gaps to meet Ubers unique scale and requirements. We then highlight several important use cases and show their real-time solutions and tradeoffs. Finally, we reflect on the lessons we learned as we built, operated and scaled these systems.
سجل دخول لتتمكن من نشر تعليقات
التعليقات
جاري جلب التعليقات
سجل دخول لتتمكن من متابعة معايير البحث التي قمت باختيارها
