Introduction
Trust is a little word that has a lot of mass. My father taught me to never trust someone that said, “Trust me”. Ronald Reagan taught us to “Trust but verify”. Both apply any time we are relying on another intelligence, artificial or real, to provide us with guidance. Blind trust is dangerous. We always need to be able to question and qualify the “opinions” of others.
The watsonx team uses the phrase “Trustworthy AI” – Artificial Intelligence worthy of our trust. Trustworthy AI requires responsible, transparent, and explainable data, AI policies and workflows. We shouldn’t trust AI just because it is intelligent - we need to be able to verify how the AI system got its answer.
watsonx is comprised of three major components that provide the foundation of many domain-specific solutions in areas of HR, marketing, product documentation, security, software development and code generation, sustainability, AIOPs, etc., with new ones showing up all the time.
In layman’s terms:
- watsonx.ai – build, train, tune, and manage AI models.
- watsonx.governance – toolkit to mitigate risks, manage regulatory requirements and address ethical concerns i.e. help create and manage trust
- watsonx.data – Fit-for-purpose data store optimized for governed data and AI workloads, supported by querying, governance, and open data formats to access and share data.
Combined, these three components provide a “Trusted AI Platform”.
One of the key components of watsonx.data is the IBM Data Fabric, an architecture that enables organizations to integrate and manage data in various formats from multiple sources through multiple protocols providing a unified view of data across the enterprise. It delivers trusted data on a scalable foundation, accelerating the impact of AI and coordinating data access across the organization.
watsonx leverages IBM Cloud Pak for Data (CP4D) to deliver the IBM Data Fabric and provide a “Trusted Data Platform” to build Trustworthy AI. CP4D is a cloud-native platform that helps unify the collection, organization, and analysis of data, allowing enterprises to turn data into value. It provides an integrated catalog of IBM, open-source, and third-party microservices, making it extensible and customizable to unique client data and AI landscapes.
The Data Fabric needs scalable and resilient infrastructure to store and retrieve data quickly and safely. watsonx applications and working data need to be protected from human and infrastructure failures. You need a “Trusted Storage Platform” to provide non-volatile storage, data and application backup, and automated disaster recovery so you can promote AI projects from development to mission critical applications.
IBM Trusted Storage Platform
Trust in the storage platform is different – and far less nuanced – compared to trust in the AI Platform. It’s a black and white proposition – the storage platform must protect the data and applications, enable recovery when something goes wrong, and meet the scalability and performance requirements of AI workloads. The AI data requirement is like one of those portals in the movies, where everything gets sucked in: structured data from databases, marketing collateral, data from SaaS platforms, historical file and object stores, emails, real-time sensor data, CSAT surveys, voice and video streams, etc. I doubt there is a type of data that a data scientist isn’t looking to mine for business value. Simply put, if the storage platform can’t fuel the AI momentum safely and securely, it can’t be trusted.
Trustworthy AI
No single storage system can meet the wide variety of requirements of the AI landscape. The IBM Trusted Storage Platform for watsonx consists of three solutions targeted at different storage workload characteristics. These can be easily, flexibly mixed and matched together to meet the varied requirements of today’s AI projects.
A purpose-built data & storage management and application protection solution suite for Red Hat OpenShift (watsonx runs on OpenShift). Fusion was designed to be managed by OpenShift administrators and provides self-service provisioning of storage resources accelerating application development and deployment.
- Persistent file, block, and object services tightly integrated with OpenShift to meet all the storage requirements for watsonx and CP4D operations.
- Object gateway providing programmatic access to S3 compatible data stores with local caching.
- Integrated application aware backup with automated recovery to protect the watsonx operational components via backup recipes.
- Data replication with automated application failover and failback.
Addresses the performance and capacity demands of high-performance computing and AI workloads. It is designed to optimize workloads and reduce data warehousing costs making it a powerful tool for AI and data-intensive projects. In addition to high-performance local storage, Scale provides a high-performance global data cache allowing watsonx projects to incorporate file and S3 data from “anywhere” without having to perform Extract-Transform-Load (ETL) operations which can significantly handicap AI projects.
- Massive performance (310 GB/s and 16M IOPs per node) and scalability (1000+ nodes, YB+ capacity) to meet the biggest data challenges and business opportunities.
- Global, high-performance unstructured data (S3 and file) cache to gain business value from the data where it resides today.
- Implemented as an appliance to minimize implementation time and support effort.
- Presto and Spark support
- Replacement for HDFS in your Hadoop modernization strategy
- Automated and transparent tiering to lower cost storage media
- File auditing and intrusion analysis and detection.
- Tested and qualified with IBM Netezza and IBM Db2 Warehouse
- FIPs 140-2 accredited encryption for sensitive data.
An open source, enterprise-grade, distributed object store providing a scalable and flexible, software-defined storage solution for the large amounts of data used in AI applications.
- Scalable, self-healing infrastructure, with the ability to be optimized for performance and / or cost to match AI workload requirements.
- Combine with watsonx.data and IBM Cloud Pak for Data to create a modern data lakehouse to simplify data access and analytical processing.
- Support for S3 Select to improve data operations efficiency by executing queries on the Ceph cluster returning only the requested parts of the object.
- Presto and Iceberg support.
- Datacenter-Data-Delivery Network (D3N) providing high-performance local cache of S3 objects.
- S3A protocol support for Spark and legacy Hadoop modernization strategy.
- FIPs 140-2 accredited encryption for sensitive data.
AI projects are data intensive by design. Generative AI projects are even more data intensive. Instead of using SME knowledge to train models, we are presenting huge amounts of data to the systems so they can teach themselves. In my layman’s view, instead of sending the AI system to class, we are pointing it at the library and telling it to teach itself. Experts have shown that it is much more efficient to point smart AI systems at mass amounts of data than for us to teach them. But - we must be careful because the system may learn the wrong things like getting confused between the fiction and non-fiction sections of the library.
A couple of keys to success – first we need to make sure the books, videos, papers, data streams, etc., in the library can be read as fast as the AI system can consume them. This the value of the IBM Trusted Storage Platform. Then we need to make sure the system learns from the right information, i.e. if the system is looking to provide a foundation for understanding writing styles over the decades, access to the fiction section is fine. However, if building a model to support historical analysis, fiction may not be a great training medium. This is the value of The Trusted Data Platform and watsonx.governance.
What does this look like in your datacenter? If you believe the GenAI proponents, this is going to be like your favorite cookie, you can never have just one. Once you realize the value of watsonx, you are going to have multiple AI projects and projects in multiple business units in your company. Many of these projects will be using common data, though some projects will be limited to portions of the ‘library’. So sooner or later, it should look like this
You will end up with multiple watsonx projects using different components of the Trusted AI Platform to address different business opportunities, all sharing the same Trusted Storage Platform
A few key points:
- Storage for watsonx is not an all-or-nothing proposition. Different watsonx use cases have different storage requirements. If you look carefully at the lists above, you will see that there is an overlap between the capabilities of IBM Scale and IBM Ceph. When you dig deeper into the technologies, one or the other may be a better fit for a specific project. The IBM watsonx SMEs work with IBM Storage SMEs to design the right fit-to-use solution for each of your problem-spaces all leveraging the same storage platform.
- This is designed to be a future-proof, extensible architecture. Each of the components supports massive scalability. So, if your third project team says they need high-speed access to massive amounts of historical data in S3 buckets on a public cloud provider, you can scale out the IBM AFM global cache. (BTW any other project that wants access to the same data can access through the same cache.) If the fourth project team decides it makes sense to bring some of that data into a local data lakehouse, you can easily scale the IBM Storage Ceph infrastructure.
- This is a very simple diagram, showing basic relationships. All data in the “library” is available to all systems with the Data Fabric provided by CP4D guiding access to correct “sections” and blocking access to inappropriate material. Reality is different projects will have different requirements; some may only use a small set of the features available. For instance, imagine one of the watsonx projects is Hadoop modernization leveraging the Scale file system to replace HDFS, or another project is leveraging S3 Select to IBM Storage Ceph to retrieve small parts of very large S3 objects. If you look at the list of key features above, you can see there are lots of possibilities. Maybe in upcoming blogs I will drill into some specific use cases and how watsonx and the Data Platform leverage the IBM Storage Platform.
IBM Fusion HCI
Now you say you aren’t ready to build out the entire Trusted Storage Platform yet, or are looking for a simpler way to deploy watsonx, IBM has a Fusion Hyper Converged Infrastructure (HCI) solution designed for watsonx that includes:
- Red Hat OpenShift compute and storage nodes including IBM Fusion with integrated backup and application recovery.
- High Speed networking.
- Optional IBM Scale AFM caching nodes for accelerating file and object data.
- Optional GPU enabled compute nodes for accelerating AI/ML workloads.
This is a great solution if you are:
- New to Red Hat OpenShift and would like to shortcut the infrastructure setup processes.
- Want to focus on getting value out of watsonx as quickly as possible.
- Need an environment with limited IT support.
- Looking for a HW/SW solution with a combined support model.
Conclusion
Trust is the key word in leveraging AI in our businesses. We need to trust the data, the models, the results, etc. Trust is achieved through building on trusted components like the IBM watsonx Trusted AI Platform and Trusted Data Platform. You also need to build on a Trusted Storage Platform that will deliver performance, scalability, and resilience not just for today but well into the future as you realize the value of the next great evolutionary step in computing helping people.
Want to hear more - check out the FusionCast on the Trusted Data Platform
Special Thanks to Craig Wilcox, Rakin Haque, Matthew Klos, and Randy Arseneau for their valuable time and ‘trusted’ opinions.
If you would like to talk to other Fusion users and / or Fusion SMEs, join the Monthly User Panel.