Software Defined Storage, Big Data and Ceph - What Is all the Fuss About? By: Kamesh Pemmaraju,Neil Levine
Have you heard about Inktank Ceph and are interested to learn some tips and tricks for getting started quickly and efficiently with Ceph? Then this is the session for you! In this two part session you learn details of: • the very latest enhancements and capabilities delivered in Inktank Ceph Enterprise such as a new erasure coded storage back-end, support for tiering, and the introduction of user quotas. • best practices, lessons learned and architecture considerations founded in real customer deployments of Dell and Inktank Ceph solutions that will help accelerate your Ceph deployment.
Injustice - Developers Among Us (SciFiDevCon 2024)
Software Defined Storage, Big Data and Ceph - What Is all the Fuss About?
1. Software Defined storage, Big
Data and Ceph.
What is all the fuss about?
Kamesh Pemmaraju, Sr. Product Mgr, Dell
Neil Levine, Dir. of Product Mgmt, Red Hat
OpenStack Summit Atlanta,
May 2014
24. Outline
• Planning your Ceph implementation
• Choosing targets for Ceph deployments
• Reference Architecture Considerations
• Dell Reference Configurations
• Customer Case Study
25. • Business Requirements
– Budget considerations, organizational commitment
– Avoiding lock-in – use open source and industry standards
– Enterprise IT use cases
– Cloud applications/XaaS use cases for massive-scale, cost-effective storage
– Steady-state vs. Spike data usage
• Sizing requirements
– What is the initial storage capacity?
– What is the expected growth rate?
• Workload requirements
– Does the workload need high performance or it is more capacity focused?
– What are IOPS/Throughput requirements?
– What type of data will be stored?
– Ephemeral vs. persistent data, Object, Block, File?
Planning your Ceph Implementation
26. How to Choose Targets Use Cases for Ceph
Virtualization and Private
Cloud
(traditional SAN/NAS)
High Performance
(traditional SAN)
PerformanceCapacity
NAS & Object
Content Store
(traditional NAS)
Cloud
Applications
Traditional IT
XaaS Compute Cloud
Open Source Block
XaaS Content Store
Open Source NAS/Object
Ceph
Target
Ceph Target
27. • Tradeoff between Cost vs. Reliability (use-case dependent)
• Use the Crush configs to map out your failures domains and performance pools
• Failure domains
– Disk (OSD and OS)
– SSD journals
– Node
– Rack
– Site (replication at the RADOS level, Block replication, consider latencies)
• Storage pools
– SSD pool for higher performance
– Capacity pool
• Plan for failure domains of the monitor nodes
• Consider failure replacement scenarios, lowered redundancies, and performance
impacts
Architectural considerations – Redundancy and
replication considerations
28. Server Considerations
• Storage Node:
– one OSD per HDD, 1 – 2 GB ram, and 1 Gz/core/OSD,
– SSD’s for journaling and for using the tiering feature in Firefly
– Erasure coding will increase useable capacity at the expense of additional compute
load
– SAS JBOD expanders for extra capacity (beware of extra latency and
oversubscribed SAS lanes)
• Monitor nodes (MON): odd number for quorum, services
can be hosted on the storage node for smaller
deployments, but will need dedicated nodes larger
installations
• Dedicated RADOS Gateway nodes for large object store
deployments and for federated gateways for multi-site
29. Networking Considerations
• Dedicated or Shared network
– Be sure to involve the networking and security teams early when design your
networking options
– Network redundancy considerations
– Dedicated client and OSD networks
– VLAN’s vs. Dedicated switches
– 1 Gbs vs 10 Gbs vs 40 Gbs!
• Networking design
– Spine and Leaf
– Multi-rack
– Core fabric connectivity
– WAN connectivity and latency issues for multi-site deployments
30. Ceph additions coming to the Dell Red Hat
OpenStack solution
Pilot configuration Components
• Dell PowerEdge R620/R720/R720XD Servers
• Dell Networking S4810/S55 Switches, 10GB
• Red Hat Enterprise Linux OpenStack Platform
• Dell ProSupport
• Dell Professional Services
• Avail. w/wo High Availability
Specs at a glance
• Node 1: Red Hat Openstack Manager
• Node 2: OpenStack Controller (2 additional controllers
for HA)
• Nodes 3-8: OpenStack Nova Compute
• Nodes: 9-11: Ceph 12x3 TB raw storage
• Network Switches: Dell Networking S4810/S55
• Supports ~ 170-228 virtual machines
Benefits
• Rapid on-ramp to OpenStack cloud
• Scale up, modular compute and storage blocks
• Single point of contact for solution support
• Enterprise-grade OpenStack software package
Storage
bundles
31. Example Ceph Dell Server Configurations
Type Size Components
Performance 20 TB • R720XD
• 24 GB DRAM
• 10 X 4 TB HDD (data drives)
• 2 X 300 GB SSD (journal)
Capacity 44TB /
105 TB*
• R720XD
• 64 GB DRAM
• 10 X 4 TB HDD (data drives)
• 2 X 300 GB SSH (journal)
• MD1200
• 12 X 4 TB HHD (data drives)
Extra Capacity 144 TB /
240 TB*
• R720XD
• 128 GB DRAM
• 12 X 4 TB HDD (data drives)
• MD3060e (JBOD)
• 60 X 4 TB HHD (data drives)
32. • Dell & Red Hat & Inktank have partnered to bring a complete
Enterprise-grade storage solution for RHEL-OSP + Ceph
• The joint solution provides:
– Co-engineered and validated Reference Architecture
– Pre-configured storage bundles optimized for performance or
storage
– Storage enhancements to existing OpenStack Bundles
– Certification against RHEL-OSP
– Professional Services, Support, and Training
› Collaborative Support for Dell hardware customers
› Deployment services & tools
What Are We Doing To Enable?
34. Overcoming a data deluge
Inconsistent data management across research teams hampers productivity
• Growing data sets challenged available resources
• Research data distributed across laptops,
USB drives, local servers, HPC clusters
• Transferring datasets to HPC clusters took too
much time and clogged shared networks
• Distributed data management reduced
researcher productivity and put data at risk
35. Solution: a storage cloud
Centralized storage cloud based on OpenStack and Ceph
• Flexible, fully open-source infrastructure
based on Dell reference design
− OpenStack, Crowbar and Ceph
− Standard PowerEdge servers and storage
− 400+ TBs at less than 41¢ per gigabyte
• Distributed scale-out storage provisions
capacity from a massive common pool
− Scalable to 5 petabytes
• Data migration to and from HPC clusters via
dedicated 10Gb Ethernet fabric
• Easily extendable framework for developing
and hosting additional services
− Simplified backup service now enabled
“We’ve made it possible for users to
satisfy their own storage needs with
the Dell private cloud, so that their
research is not hampered by IT.”
David L. Shealy, PhD
Faculty Director, Research Computing
Chairman, Dept. of Physics
36. Building a research cloud
Project goals extend well beyond data management
• Designed to support emerging
data-intensive scientific computing paradigm
– 12 x 16-core compute nodes
– 1 TB RAM, 420 TBs storage
– 36 TBs storage attached to each compute node
• Virtual servers and virtual storage meet HPC
− Direct user control over all aspects of the
application environment
− Ample capacity for large research data sets
• Individually customized test/development/
production environments
− Rapid setup and teardown
• Growing set of cloud-based tools & services
− Easily integrate shareware, open source, and
commercial software
“We envision the OpenStack-based
cloud to act as the gateway to our
HPC resources, not only as the
purveyor of services we provide, but
also enabling users to build their own
cloud-based services.”
John-Paul Robinson, System Architect
37. Research Computing System (Next Gen)
A cloud-based computing environment with high speed access to
dedicated and dynamic compute resources
Open
Stack
node
Open
Stack
node
Open
Stack
node
Open
Stack
node
Open
Stack
node
Open
Stack
node
Open
Stack
node
Open
Stack
node
Open
Stack
node
Open
Stack
node
Open
Stack
node
Open
Stack
node
HPC
Cluster
HPC
Cluster
HPC
Storage
DDR Infiniband QDR Infiniband
10Gb Ethernet
Cloud services layer
Virtualized server and storage computing cloud
based on OpenStack, Crowbar and Ceph
UAB Research Network
39. Contact Information
Reach Kamesh and Neil for additional information:
Dell.com/OpenStack
Dell.com/Crowbar
Inktank.com/Dell
Kamesh_Pemmaraju@Dell.com
@kpemmaraju
Neil.Levine@Inktank.com
@neilwlevine
Visit the Dell and Inktank booths in the OpenStack Summit Expo Hall
Editor's Notes
R720XD configurations use
4 TB drives
2 X 300 GB OS drives
2 X 10 GB NIC
iDRAC 7 Enterprise
LSI 9207-[8i, 8e] HBAs
2 X E5-2650 2 GHz processors
(*) - The larger capacity is that were erasure encoding is in use. To get the same redundancy as 2 X in erasure encoding uses a factor of 1.2. Erasure encoding is a feature of the Ceph Firefly release, which is in its final phase of development.
Additional performance could be gained by adding either Intel’s CAS or Dell FluidFS DAS caching software packages. Doing so would impose additional memory and processing overhead, and more work in the deployment/installation bucket (because we would have to install and configure it).
https://dev.uabgrid.uab.edu/wiki/OpenStackPlusCeph
The research computing system (RCS) is built on a collection of distinct hardware systems designed to provide specific services to applications. The RCS hardware includes dedicated compute fabrics that support high performance computing (HPC) applications where hundreds of compute cores can work together on a single application. These clusters of commodity compute hardware make it possible to do data analysis and modelling work in hours, work that would have taken months using a single computer. The clusters are connected with dedicated high bandwidth, low latency networks for applications to efficiently coordinate their actions across many computers and access a shared high speed storage system for working efficiently with terabytes of data.
Our newest hardware fabric, acquired 2012Q4, is designed to support emerging data intensive scientific computing and virtualization paradigms. This hardware is very similar to the commodity computers used by our traditional HPC fabrics, however, in addition to having many compute cores and lots of RAM, each individual computer contains 36TB of built in disk storage. Taken together, this newest hardware fabric adds 192 cores, 1TB RAM, and 420TB of storage to the RCS.
The built-in disk storage is designed to support applications running local to each computer. The data intensive computing paradigm exchanges the external storage networks of traditional HPC clusters with the native, very high speed system buses that provide access to local hard disks in each computer. Large datasets are distributed across these computers and then applications are assigned to run on the specific computer that stores the portion of the dataset it has been assigned to analyze. The hardware requirements for data intensive computing closely resemble the requirements for virtualization and can benefit tremendously from the configuration flexibility that a virtualization fabric offers.
In order to enhance flexibility and further improve support for scaling research applications, we are engineering our latest hardware cluster to act as a virtualized storage and compute fabric. This enables support for a wide variety of storage and compute use cases, most prominently, ample storage capacity for reliably housing large research data collections and flexible application development and deployment capabilities that allow direct user control over all aspects of the application environment.
In short, we are tooling this hardware to build a cloud computing environment.
We are building this cloud using OpenStack for compute virtualization and Ceph for storage virtualization. Crowbar will provision the raw hardware fabric. This approach is very similar to the mode we have been following with our traditional ROCKS-based HPC cluster environment. The new approach enhances our ability to automatically provision hardware and further improve the economics large scale computing.
We are implementing this environment with Dell and Inktank. These vendors and the upstream open source projects on which this platform is built, embrace the DevOps model for systems development. This will support further engineering collaboration with our vendors, enabling the UAB research community to continually enhance our fabric as needed and feed those enhancements upstream for inclusion in future support releases.
This solution rounds out the feature set of the RCS core and will provide a general framework to scale future growth.
https://dev.uabgrid.uab.edu/wiki/OpenStackPlusCeph
The research computing system (RCS) is built on a collection of distinct hardware systems designed to provide specific services to applications. The RCS hardware includes dedicated compute fabrics that support high performance computing (HPC) applications where hundreds of compute cores can work together on a single application. These clusters of commodity compute hardware make it possible to do data analysis and modelling work in hours, work that would have taken months using a single computer. The clusters are connected with dedicated high bandwidth, low latency networks for applications to efficiently coordinate their actions across many computers and access a shared high speed storage system for working efficiently with terabytes of data.
Our newest hardware fabric, acquired 2012Q4, is designed to support emerging data intensive scientific computing and virtualization paradigms. This hardware is very similar to the commodity computers used by our traditional HPC fabrics, however, in addition to having many compute cores and lots of RAM, each individual computer contains 36TB of built in disk storage. Taken together, this newest hardware fabric adds 192 cores, 1TB RAM, and 420TB of storage to the RCS.
The built-in disk storage is designed to support applications running local to each computer. The data intensive computing paradigm exchanges the external storage networks of traditional HPC clusters with the native, very high speed system buses that provide access to local hard disks in each computer. Large datasets are distributed across these computers and then applications are assigned to run on the specific computer that stores the portion of the dataset it has been assigned to analyze. The hardware requirements for data intensive computing closely resemble the requirements for virtualization and can benefit tremendously from the configuration flexibility that a virtualization fabric offers.
In order to enhance flexibility and further improve support for scaling research applications, we are engineering our latest hardware cluster to act as a virtualized storage and compute fabric. This enables support for a wide variety of storage and compute use cases, most prominently, ample storage capacity for reliably housing large research data collections and flexible application development and deployment capabilities that allow direct user control over all aspects of the application environment.
In short, we are tooling this hardware to build a cloud computing environment.
We are building this cloud using OpenStack for compute virtualization and Ceph for storage virtualization. Crowbar will provision the raw hardware fabric. This approach is very similar to the mode we have been following with our traditional ROCKS-based HPC cluster environment. The new approach enhances our ability to automatically provision hardware and further improve the economics large scale computing.
We are implementing this environment with Dell and Inktank. These vendors and the upstream open source projects on which this platform is built, embrace the DevOps model for systems development. This will support further engineering collaboration with our vendors, enabling the UAB research community to continually enhance our fabric as needed and feed those enhancements upstream for inclusion in future support releases.
This solution rounds out the feature set of the RCS core and will provide a general framework to scale future growth.
https://dev.uabgrid.uab.edu/wiki/OpenStackPlusCeph
The research computing system (RCS) is built on a collection of distinct hardware systems designed to provide specific services to applications. The RCS hardware includes dedicated compute fabrics that support high performance computing (HPC) applications where hundreds of compute cores can work together on a single application. These clusters of commodity compute hardware make it possible to do data analysis and modelling work in hours, work that would have taken months using a single computer. The clusters are connected with dedicated high bandwidth, low latency networks for applications to efficiently coordinate their actions across many computers and access a shared high speed storage system for working efficiently with terabytes of data.
Our newest hardware fabric, acquired 2012Q4, is designed to support emerging data intensive scientific computing and virtualization paradigms. This hardware is very similar to the commodity computers used by our traditional HPC fabrics, however, in addition to having many compute cores and lots of RAM, each individual computer contains 36TB of built in disk storage. Taken together, this newest hardware fabric adds 192 cores, 1TB RAM, and 420TB of storage to the RCS.
The built-in disk storage is designed to support applications running local to each computer. The data intensive computing paradigm exchanges the external storage networks of traditional HPC clusters with the native, very high speed system buses that provide access to local hard disks in each computer. Large datasets are distributed across these computers and then applications are assigned to run on the specific computer that stores the portion of the dataset it has been assigned to analyze. The hardware requirements for data intensive computing closely resemble the requirements for virtualization and can benefit tremendously from the configuration flexibility that a virtualization fabric offers.
In order to enhance flexibility and further improve support for scaling research applications, we are engineering our latest hardware cluster to act as a virtualized storage and compute fabric. This enables support for a wide variety of storage and compute use cases, most prominently, ample storage capacity for reliably housing large research data collections and flexible application development and deployment capabilities that allow direct user control over all aspects of the application environment.
In short, we are tooling this hardware to build a cloud computing environment.
We are building this cloud using OpenStack for compute virtualization and Ceph for storage virtualization. Crowbar will provision the raw hardware fabric. This approach is very similar to the mode we have been following with our traditional ROCKS-based HPC cluster environment. The new approach enhances our ability to automatically provision hardware and further improve the economics large scale computing.
We are implementing this environment with Dell and Inktank. These vendors and the upstream open source projects on which this platform is built, embrace the DevOps model for systems development. This will support further engineering collaboration with our vendors, enabling the UAB research community to continually enhance our fabric as needed and feed those enhancements upstream for inclusion in future support releases.
This solution rounds out the feature set of the RCS core and will provide a general framework to scale future growth.
User base: 900+ researchers across Campus.
KVM-based
2 Nova nodes
4 primary storage nodes
4 replication nodes
2 control nodes
12 x R720XD systems