MemVerge: Gismo (Global IO-free Shared Memory Objects)

•Download as PPTX, PDF•

0 likes•299 views

Project Gismo introduces a global I/O-free shared memory object (Gismo) library that utilizes CXL to provide direct memory access across nodes. This allows distributed applications to access remote objects as fast as local memory, eliminating object serialization and data copying. Demo results show Gismo can improve performance of AI/ML workloads like Ray by up to 675% and reduce database synchronization times. The Gismo API provides functions to connect, create, access, and manage shared memory objects globally without I/O.

Technology

Project Gismo
Global I/O-free Shared Memory Objects

The I/O Wall in Distributed Applications
2
Node
App
DRAM
Node
App
DRAM
Storage
Node
App
DRAM
Node
App
DRAM
Network
Serialization &
Deserialization
Media Speed
Data Copies
MemVerge Inc. Confidential

CXL Makes Cross-node Shared Memory Possible!
3
CXL Shared Memory
Node
App
DRAM
Node
App
DRAM
• Multiple nodes have direct
memory access to the same
CXL memory region
• Cache Coherence is part of
CXL 3.0 hardware
specification
• Software cache coherence
can be implemented on top of
CXL 2.0 hardware
MemVerge Inc. Confidential

Introducing Project Gismo
Global I/O-free Shared Memory Objects
4
Node
App
Gismo Library
Gismo
Manager
CPU
Node
App
Gismo Library
CPU
Node
App
Gismo Library
CPU
Shared Memory over CXL
DDR DRAM
(NUMA 0)
DDR DRAM
(NUMA 0)
DDR DRAM
(NUMA 0)
MemVerge Inc. Confidential

5
Baseline Ray
Ray+Gismo
Use Case 1: AI/ML
MemVerge Inc. Confidential

Baseline Ray Ray + Gismo
Local Get 1GB object 0.4 sec
0.4 sec CXL shared memory
as fast as local memory
Remote Get 1GB object 2.7 sec 0.4 sec 675% faster
Shuffle 50GB
4 nodes, each 4 cores,
128 GB object store
515 sec 185 sec 280% faster
* Running in emulation environment
Shuffle Benchmark Results
7
MemVerge Inc. Confidential

Benefits of Ray+Gismo
● IO-free: No more object
serialization and transfer over
network for remote object access
● Zero Copy: No more duplicate
object copies on different nodes
● No Spilling: Reduce object
spilling and data skewing
because each node has access
to the whole memory pool
8
MemVerge Inc. Confidential

Use Case 2: Memory Cache for HTAP Database
Transaction Engine
S3
Transaction Engine Transaction Engine
Local Cache Local Cache Local Cache
• MatrixOrigin HTAP database features a scale-out architecture
• Local cache can have high cache miss rate
• Cache coherence is complicated.
MemVerge Inc. Confidential

Gismo Enables Shared Memory Cache
Transaction Engine
S3
Transaction Engine Transaction Engine
Gismo Shared
Memory Cache
• Higher cache hit rate
• Less cache copies
• No cache coherency handling
MemVerge Inc. Confidential

Use Case 3: Database Fail-over
11
• Timeplus distributed streaming analytics used for real-time trade monitoring
• Synchronization of data for fail-over recovery is time consuming
Node Node
1. Query Data
2. Query Index
3. Query State
SSD
1. Query Data
2. Query Index
3. Query State
SSD
MemVerge Inc. Confidential

Faster Database Failover with Gismo
• Using Gismo-based checkpoint and restore
• Synchronization of data and fail-over recovery will be significantly faster
Node Node
1. Query Data
2. Query Index
3. Query State
1. Query Data
2. Query Index
3. Query State
Gismo Gismo
Shared Memory over CXL
MemVerge Inc. Confidential

Gismo_connect(): Connect to Gismo Manager
Gismo_disconnect(): Disconnect from Gismo Manager
Gismo_create(): Create an object
Gismo_seal(): Seal an object. Object becomes immutable after this call
Gismo_get(): get/read an object
Gismo_release(): Tell the Gismo manager that this host is done reading the object
Gismo_delete(): Delete an object
Gismo_subscribe(): Subscribe to notifications when objects are sealed in Gismo
Gismo API
13
MemVerge Inc. Confidential

14
CXL+MemVerge
Memory is the Network
Memory is the Storage
Welcome to Memory-Centric Computing
MEMVERGE CONFIDENTIAL

15
Contact gismo@memverge.com
● If you are a developer of distributed applications
● If you are in the field of AI/ML or distributed databses
● If you are looking for differentiated performance
advantages for your applications
MemVerge Inc. Confidential

MemVerge: Gismo (Global IO-free Shared Memory Objects)

Customers regularly use Apache Spark running on Amazon EMR to process large amounts of data. As time to insight and the ability to act quickly based on those insights become core differentiators for customers, there is a greater need to be able to analyze data in real time. In this session, we teach you several design patterns to process and analyze real-time streaming data using Amazon EMR and Amazon Kinesis data services.

Getting started with postgresql

botsplash.com

This document provides an overview of Postgresql, including its history, capabilities, advantages over other databases, best practices, and references for further learning. Postgresql is an open source relational database management system that has been in development for over 30 years. It offers rich SQL support, high performance, ACID transactions, and extensive extensibility through features like JSON, XML, and programming languages.

Reducing Database Pain & Costs with Postgres

EDB

View this presentation to gain insight into optimizing Postgres and savings for your data management. Visit EntepriseDB's > Resources > Webcasts to view the presentation by Jay Barrows, VP of Field Operations. During this 45 -minute presentation, Jay Barrows, VP of Field Operations, will provide a business review of how, where and why businesses are leveraging PostgreSQL. In addition, he will go over the primary pains and business drivers shaping the data management landscape such as significant cost pressures combined with recent improvements to open source database options. Oracle migration is often considered the most powerful cost reduction opportunity if you understand the migration risks, and have a clear migration game plan. Jay will discuss several use cases selected that highlight how enterprise customers are leveraging their findings from the adoption of other OSS products, to helping to bring Postgres to the extremely expensive and mission critical part of their IT stack - the DB. By doing so they are driving TCO down in very meaningful ways, sacrificing nothing in terms of performance, scalability, security or reliability. Many businesses are already leveraging OSS in much lower cost parts of IT stack (OS, middleware). This presentation will be beneficial to decision-makers interested in enhancing their data management with PostgreSQL. I

Greenplum 6 Changes

VMware Tanzu

This document summarizes new features in Greenplum 6 including changes to querying, data types, storage, extensions, administration, backup/restore, security, and other areas. Some key changes are: - Support for collations to specify sort order at the column level - Enhanced recursive queries including support for cycles and path tracking - Improved error handling for stored procedures - Introduction of the JSONB data type for more efficient JSON storage and querying - Support for replicated tables to replicate data across all segments - New compression algorithms including Zstandard - Enhancements to the CREATE TABLE statement including LIKE support to copy properties

InfluxDB IOx Tech Talks: Query Processing in InfluxDB IOx

InfluxData

PostgreSQL HA

haroonm

New Generation Oracle RAC 19c focuses on diagnosing Oracle RAC performance issues. The document discusses tools used by Oracle's RAC performance engineering team to instrument and measure key code areas between releases. It also covers how Oracle RAC provides high availability and scalability for workloads like traditional apps, new apps, IoT workloads, and more. Diagnosing performance requires understanding factors like private network latency and configuration.

Presto anatomy

Dongmin Yu

Presto is an open source distributed SQL query engine for running interactive analytic queries against data sources of all sizes ranging from gigabytes to petabytes. It is written in Java and uses a pluggable backend. Presto is fast due to code generation and runtime compilation techniques. It provides a library and framework for building distributed services and fast Java collections. Plugins allow Presto to connect to different data sources like Hive, Cassandra, MongoDB and more.

Lessons for the optimizer from running the TPC-DS benchmark

Sergey Petrunya

Practical Memory Tuning for PostgreSQL

Grant McAlister

This talk will first introduce the different ways PostgreSQL can use memory, from the operating system, to cluster wide and then into per session and per operation. From there we will dive into specifics around different PostgreSQL parameters like shared_buffers, work_mem, maintenance_work_mem and how to set them depending on your workload. The presentation will also cover some of the lesser known ways that PostgreSQL will consume memory, how you can diagnose what is using the memory in your PostgreSQL cluster and possible ways to avoid running out of memory. Additionally the talk we will cover the importance of hugepages for not only performance but memory usage on large memory systems.

Canadian CNCF: "Emissary-ingress 101: An introduction to the CNCF incubation-...

Daniel Bryant

By Daniel Bryant, Ambassador Labs We all need to be able to get user traffic into our applications, and your requirements for services running on Kubernetes are no different. "But", I hear you say, "what about the K8s Ingress spec? And how do I observe what's happening under the hood? And who should be responsible for configuring the gateway: dev or ops?" These are all good questions! Join me for a whistle-stop tour of all things emissary-ingress, where we will explore how this new edition to the family of CNCF incubation projects can make your life easier when it comes to routing, observability, and integration into the bigger (people and technology) picture.

Using Angular JS in APEX

Enkitec

This document discusses using the AngularJS framework within Oracle Application Express (APEX) applications. It provides background on the presenter and his company, which focuses on Oracle technologies. It then explains some key aspects of integrating AngularJS and APEX, such as linking Angular asset files, modifying the page template, and standardizing Ajax processes. While AngularJS is powerful and feature-rich, it also has a steep learning curve and is more complex than standard APEX development.

MongoDB + Java + Spring Data

Anton Sulzhenko

This document provides an overview of MongoDB, Java, and Spring Data. It discusses how MongoDB is a document-oriented NoSQL database that uses JSON-like documents with dynamic schemas. It describes how the Java driver can be used to interact with MongoDB to perform CRUD operations. It also explains how Spring Data provides an abstraction layer over the Java driver and allows for object mapping and repository-based queries to MongoDB.

Oracle to Postgres Migration - part 2

PgTraining

Oracle Office Hours - Exposing REST services with APEX and ORDS

Doug Gault

This document provides an overview and agenda for exposing RESTful services using Oracle Application Express (APEX) and Oracle REST Data Services (ORDS). It begins with a brief introduction to REST concepts and then discusses the history of supporting REST in APEX and ORDS. The document outlines the key components and APIs available in ORDS for enabling schemas and defining REST modules, templates and handlers to expose database objects and custom services via REST. It also provides examples of enabling the schema and using auto-REST to expose database tables through REST.

Using LLVM to accelerate processing of data in Apache Arrow

DataWorks Summit

Most query engines follow an interpreter-based approach where a SQL query is translated into a tree of relational algebra operations then fed through a conventional tuple-based iterator model to execute the query. We will explore the overhead associated with this approach and how the performance of query execution on columnar data can be improved using run-time code generation via LLVM. Generally speaking, the best case for optimal query execution performance is a hand-written query plan that does exactly what is needed by the query for the exact same data types and format. Vectorized query processing models amortize the cost of function calls. However, research has shown that hand-written code for a given query plan has the potential to outperform the optimizations associated with a vectorized query processing model. Over the last decade, the LLVM compiler framework has seen significant development. Furthermore, the database community has realized the potential of LLVM to boost query performance by implementing JIT query compilation frameworks. With LLVM, a SQL query is translated into a portable intermediary representation (IR) which is subsequently converted into machine code for the desired target architecture. Dremio is built on top of Apache Arrow’s in-memory columnar vector format. The in-memory vectors map directly to the vector type in LLVM and that makes our job easier when writing the query processing algorithms in LLVM. We will talk about how Dremio implemented query processing logic in LLVM for some operators like FILTER and PROJECT. We will also discuss the performance benefits of LLVM-based vectorized query execution over other methods. Speaker Siddharth Teotia, Dremio, Software Engineer

使用 Apache IoTDB 构建工业时序数据管理解决方案的实践

ZhangZhengming

Admission Control in Impala

Cloudera, Inc.

This document discusses admission control in Impala to prevent oversubscription of resources from too many concurrent queries. It describes the problem of all queries taking longer when too many run at once. It then outlines Impala's solution of adding admission control by throttling incoming requests, queuing requests when workload increases, and executing queued requests when resources become available. The document provides details on how Impala implements admission control in a decentralized manner without requiring Yarn/Llama to handle throttling and queuing locally on each Impalad daemon.

Oracle AHF Insights 23c: Deeper Diagnostic Insights for your Oracle Database

Sandesh Rao

AHF Insights provides a unified reporting and correlation tool that collects diagnostic data from an AHF stack and generates an offline report. It captures information on system topology, insights like events and issues, and allows drilling down into specific components for root cause analysis. The report includes sections on the cluster summary, resources, databases, metrics, best practices, system changes, software recommendations, and detailed parameters for databases and kernels. It provides visualizations of events and metrics over time and detects operating system issues and configuration problems.

Oracle RAC 19c: Best Practices and Secret Internals

Anil Nair

Oracle Real Application Clusters 19c provides best practices and new features for upgrading to Oracle 19c. It discusses upgrading Oracle RAC to Linux 7 with minimal downtime using node draining and relocation techniques. Oracle 19c allows for upgrading the Grid Infrastructure management repository and patching faster using a new Oracle home. The presentation also covers new resource modeling for PDBs in Oracle 19c and improved Clusterware diagnostics.

Oracle 12c Multitenant architecture

naderattia

HBase in Practice

DataWorks Summit/Hadoop Summit

HBase is a distributed, column-oriented database that stores data in tables divided into rows and columns. It is optimized for random, real-time read/write access to big data. The document discusses HBase's key concepts like tables, regions, and column families. It also covers performance tuning aspects like cluster configuration, compaction strategies, and intelligent key design to spread load evenly. Different use cases are suitable for HBase depending on access patterns, such as time series data, messages, or serving random lookups and short scans from large datasets. Proper data modeling and tuning are necessary to maximize HBase's performance.

Transactional operations in Apache Hive: present and future

DataWorks Summit

Apache Hive is an enterprise data warehouse build on top of Hadoop. Hive supports insert, update, delete, and merge SQL operations with transactional semantics and read operations that run at snapshot isolation. The well defined semantics of these operations in the face of failure and concurrency are critical to building robust application on top of Apache Hive. In the past there were many preconditions to enabling these features which meant giving up other functionality. The need to make these tradeoffs is rapidly being eliminated. This talk will describe the intended use cases, architecture of the implementation, recent improvements and new features build for Hive 3.0. For example, bucketing transactional tables, while supported, is no longer required. Performance overhead of using transactional tables is nearly eliminated relative to identical non-transactional tables. We’ll also cover Streaming Ingest API, which allows writing batches of events into a Hive table without using SQL. Speaker Eugene Koifman, Hortonworks, Principal Software Engineer

Azure cosmos db, Azure no-SQL database,

BRIJESH KUMAR

MemVerge - The Dawn of Big Memory

MemVerge: Gismo (Global IO-free Shared Memory Objects)

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to MemVerge: Gismo (Global IO-free Shared Memory Objects)

Similar to MemVerge: Gismo (Global IO-free Shared Memory Objects) (20)

More from Memory Fabric Forum

More from Memory Fabric Forum (20)

Recently uploaded

Recently uploaded (20)

MemVerge: Gismo (Global IO-free Shared Memory Objects)