## GPU Resource Management On JDOS 梁永清 liangyongqing1@jd.com ## 提供的服务 ## Experiment ## Training 1. 用于实验的 GPU 容器 2. 基于 Kubeflow 的机器学习训练服务 3. 模型管理和模型 Serving 服务 ## Serving 均基于容器，不对业务方直接提供 GPU 物理机

## +24 ## Multi Producer, Multi Consumer, Lock Free, Atomic Queue User API and Implementation ... ## EREZ STRAUSS ## Lockfree, MPMC Queue - Legal All Statements and representations are my own and do MPMC Queue - Agenda 1. Trading Platform, Latency 2. Queues Environment 3. Queues Requirement 4. Atomic hardware and C++ Atomic 5. Queues Classification 6. Queue creation, push, pop 7. Queue internals internals 8. Testing 9. Performance - Benchmarking ## Lockfree, MPMC Queue - Background Trading platform - every message is tracked, at microseconds resolution - Queues are essential data structure used

# 动态资源管理和容器技术 在金融行业的架构探索和明天 ## Resource Scheduling & Container Technology for Financial Service 余军 Gopher China 2015 ## 关于我 ☐ ~19y+ Linux & opensource experience ☐ Co-founder of Shanghai p14_2.jpg) ## Algorithm 1 DRF pseudo-code $R = \langle r_{1}, \cdots, r_{m} \rangle$ ▷ total resource capacities $C = \langle c_{1}, \cdots, c_{m} \rangle$ ▷ consumed resources, initially 0 $s_{i}$ 采用了DRF(Dominant Resource Fairness) 调度机制。YARN自带FIFO、Capacity Scheduler和Fair Scheduler(借鉴了Mesos的DRF)。 ② Mesos中的DRF调度算法过分的追求公平，没有考虑到实际的应用需求。在实际生产线上，往往需要类似于Hadoop中Capacity Scheduler的调度机制，将所有资源分成若干个queue，每个queue分配一定量

## +21 ## Compile-Time Compression and Resource Generation with C++20 ## ASHLEY ROLL 20 21 October 24-29 ## I ntroduction Explore how C++20's constexper features can: • Generate data from code that take a user-supplied lambda to generate the data needed to render our desired compile-time resource! - These are effectively templated functions, but we will use the cleaner auto parameter syntax symbols and their frequency • symbols are all characters in all strings - Uses a min-heap priority_queue to build a tree bottom up from least used to most used - Bit pattern assigned by walking down the

## Greenplum资源管理器 姚珂男/Pivotal kyao@pivotal.io ## Agenda • Greenplum数据库 • Resource Queue • Resource Group ## Greenplum数据库 • 基于PostgreSQL • 分布式 • OLAP • MPP(Massively Parallel Processing) ## Greenplum数据库 e7818fe3eb6bc09e8312568f4d7a/p4_1.jpg) ## Resource Queue • SQL语句并发控制 • 基于cost的并发控制 • 基于priority的CPU控制 • 内存控制 ## Running Example • CREATE RESOURCE QUEUE rq WITH ( active_statements = 6, max_cost min_cost = 50000, priority = high, memory_limit = '1024MB' ); • CREATE ROLE r1 RESOURCE QUEUE rq; • SELECT * FROM gp_toolkit.gp_resqueue_status; ## 内存控制 • virtual memory note keeping (gp_malloc)

http://github.com/ask/celery/ Keywords task queue, job queue, asynchronous, rabbitmq, amqp, redis, Django, Python, Webhooks, queue, distributed Celery is a task queue/job queue based on distributed message passing RabbitMQ. You can set up new workers without central configuration (e.g. use your dads laptop while the queue is temporarily overloaded).| |Concurrency|Tasks are executed in parallel using the multiprocessing key used to deliver this task. It’s used by e.g. retry() to resend the task to the same destination queue. NOTE As some messaging backends doesn’t have advanced routing capabilities, you can’t trust the

version of Celery (2.2). For the latest stable version please go here. ## Celery - Distributed Task Queue Contents: • Getting Started • Introduction • Broker Installation • First steps com/ask/celery/ Keywords: task queue, job queue, asynchronous, rabbitmq, amqp, redis, python, webhooks, queue, distributed Celery is an open source asynchronous task queue/job queue based on distributed message set up new workers without central configuration (e.g. use your grandma's laptop to help if the queue is temporarily congested).| |Concurrency|Concurrency is achieved by using multiprocessing, Eventlet

Detecting Dead Connections 1.21 Detecting Slow Consumers 1.22 Avoiding Network Isolation 1.23 Resource Manager Configuration 1.24 Flow Control 1.25 Guarantees of sends and commits 1.26 Message 30 Queue Attributes 1.31 Scheduled Messages 1.32 Last-Value Queues 1.33 Message Grouping 1.34 Extra Acknowledge Modes 1.35 Management 1.36 Security 1.37 Broker Plugins 1.38 Resource Limits asynchronous messaging: message queue messaging (also known as point-to-point messaging) and publish subscribe messaging. We'll summarise them briefly here: ## The Message Queue Pattern With this type of

Consumers 1.28 Avoiding Network Isolation 1.29 Detecting Broker Issues (Critical Analysis) 1.30 Resource Manager Configuration 1.31 Flow Control 1.32 Guarantees of sends and commits 1.33 Message Management Console 1.47 Metrics 1.48 Security 1.49 Masking Passwords 1.50 Broker Plugins 1.51 Resource Limits 1.52 The JMS Bridge 1.53 Client Reconnection and Session Reattachment 1.54 Diverting Duplicate Message Detection 1.58 Clusters 1.59 Federation 1.60 Address Federation 1.60.1 Queue Federation 1.60.2 High Availability and Failover 1.61 Graceful Server Shutdown 1.62 Libaio Native

## Welcome to Django Q Django Q is a native Django task queue and worker application using Python multiprocessing. ## Features • Multiprocessing worker pool • Asynchronous tasks • Scheduled and repeated 'timeout': 60, 'compress': True, 'save_limit': 250, 'queue_limit': 500, 'cpu_affinity': 1, 'label': 'Django Q', 'redis': • Set to -1 for no success storage at all. • Defaults to 250 • Failures are always saved. ## queue_limit This does not limit the amount of tasks that can be queued overall on Redis, but rather how