## 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

# 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 $s_{i}$ \max_{j=1}^{m} \{ u_{i,j} / r_{j} \}$ else return ▷ the cluster is full end if ① Mesos 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

## +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

Zstandard Compression Algorithm 218 Relaxed Rules for Specifying Table Distribution Columns 218 Resource Groups Features 219 PL/pgSQL Procedural Language Enhancements 219 Replicated Table Data 220 Concurrency Oriented Storage 324 Compression 325 Distributions 325 Resource Queue Memory Management 325 Partitioning 327 Indexes 327 Resource Queues 327 Monitoring and Maintenance 328 ANALYZE 328 Vacuum 329 Settings 334 Number of Segments per Host 334 Resource Queue Segment Memory Configuration 334 Resource Queue Statement Memory Configuration 335 Resource Queue Spill File Configuration 336 Schema Design

Detecting Dead Connections 1.19 Detecting Slow Consumers 1.20 Avoiding Network Isolation 1.21 Resource Manager Configuration 1.22 Flow Control 1.23 Guarantees of sends and commits 1.24 Message 1.29 Scheduled Messages 1.30 Last-Value Queues 1.31 Message Grouping 1.32 Extra Acknowledge Modes 1.33 Management 1.34 Security 1.35 Resource Limits 1.36 The JMS Bridge 1.37 Client Reconnection messages, and transformation can also be hooked in. Apache ActiveMQ Artemis also allows routing between queues to be configured in server side configuration. This allows complex routing networks to be set up

## 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 6d5e7818fe3eb6bc09e8312568f4d7a/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, 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

range accepts EHCache 2.x as well * added "contentLength()" method to Resource abstraction * URL-based Resource variants introspect "last-modified" and "content-length" response collections and nested collections (fixed regression) * ConversionService properly handles nested Resource arrays in Map values (fixed regression) * ConversionService does not accidentally use copy constructor any value type * SpringTemplateLoader for FreeMarker supports last-modified timestamp through Resource abstraction * HibernateJpaDialect correctly closes borrowed connections even for nested JDBC executions

what is there to Rethink? ## Rethinking: Task Queues  ## Problem #1 - Task Queues Do Not Scale Well: • Contention: ☐ Even Even the most meticulously designed lock-free queues experience a significant drop in performance as the number of threads increases. - Multiple sub- queues can be used to mitigate contention but this introduces ## Rethinking: Task Queues  ## Problem #2 - No Inherent Support For Prioritization: - Priority queues address this but also

Extending Flask-RESTful • Content Negotiation • Custom Fields & Inputs ○ Response Formats ☐ Resource Method Decorators Custom Error Handlers • Intermediate Usage ○ Project Structure Use With Blueprints flask_restful import Resource, Api app = Flask(__name__) api = Api(app) class HelloWorld(Resource): def get(self): return {'hello': 'world'} api.add_resource(HelloWorld, '/') defining methods on your resource. A basic CRUD resource for a todo application (of course) looks like this: from flask import Flask, request from flask_restful import Resource, Api app = Flask(__name__)

• Co-location @ Alibaba • Kubernetes vs YARN • Kubernetes & YARN: a hybrid architecture • Resource Isolation • Future ## What/Why co-location ## Data center Avg utilization ![Image](/uploads/d in co-location • Resource sharing: Make services and jobs share a single cluster and have them run on the same hardware • Resource Isolation: Rely on scheduling efficiency and resource isolation to guarantee guarantee the SLA. ## Why co-location • Two separated cluster cause huge resource wastage. Increase resource utilization. • 1 hour peak traffic on double 11 requires a lot of extra resources. However,