AI workloads in cloud environments pose unique performance challenges, particularly in managing data pipelines and processing I/O operations. These workloads demand storage solutions that deliver both low latency and high throughput while minimizing overhead, especially in virtualized environments. To meet these demands, it is crucial to provide virtual machines with high-performance block devices that deliver minimal latency and incur minimal overhead. At Xinnor, we’ve developed xiRAID Opus, a solution tailored to overcome these challenges. This blog post outlines how xiRAID Opus addresses these obstacles by delivering superior block device performance, seamless virtualization pass-through, and efficient integration with parallel file systems. This study was first presented at the SNIA SDC event in September 2024.
AI Workloads and Performance Challenges in the Cloud
One of the main performance issues in cloud environments, particularly for Software-Defined Storage (SDS), is handling the diverse workload profiles of AI applications. Different stages in an AI data pipeline can have widely varying demands. For example, some workloads require low latency and random small I/O operations, while others need high throughput for large sequential file transfers.
Traditional SDS systems are typically optimized for one workload type, making it difficult for them to perform well across different tasks. This limitation is compounded by the losses in performance introduced by virtualization, which often causes significant bottlenecks. Additionally, many SDS solutions lack high-performance shared volume support, further restricting their scalability and flexibility in distributed cloud environments.
).
Block Device: xiRAID Opus
At the core of our high-performance storage solution is xiRAID Opus, a block device optimized for cloud environments. The ability to deliver high-performance block storage is crucial for ensuring optimal performance in virtualized AI workloads. xiRAID Opus is specifically designed to provide minimal latency and minimal overhead when delivering block devices to VMs. This performance optimization is key for handling the demanding I/O operations associated with AI data pipelines.
Whether dealing with small random I/O or large sequential data transfers, xiRAID Opus consistently maintains high performance across virtual environments, ensuring seamless data access for AI applications.
Key features of xiRAID Opus include:
- Creation of RAID-protected volumes
- Provisioning of volumes to VMs
- Two major optimizations for performance enhancement:
- Polling (significantly reduces latency by actively checking for I/O completions);
- Zero-copy (eliminates unnecessary data copying within the storage pipeline, boosting throughput).
xiRAID Opus also offers significant deployment flexibility. Whether deployed in a bare-metal setup, as a virtual appliance, or on a DPU, xiRAID Opus maintains its performance characteristics across different infrastructure configurations.
Virtualization Pass-Through Method
To ensure optimal performance in virtualized environments, it is critical to provide virtual machines with high-performance block device that deliver minimal latency and overhead. Virtualization pass-through technology is the most efficient method for delivering this block device, ensuring that the performance of AI workloads remains uncompromised in cloud environments. There are multiple methods, each offering different benefits and performance trade-offs.
In our solution, we focus on three primary methods for delivering block device in virtual environments:
VIRTIO: This widely-used interface supports both single I/O threads and multiple I/O threads, allowing efficient block device delivery.
vhost-user-blk: A local interface that passes block device directly to virtual machines, operating entirely in user space. It ensures high performance by using a zero-copy approach, which reduces unnecessary data movement. At Xinnor, we've developed multithreading support for vhost-user-blk, a feature unique to our implementation. This support significantly boosts performance, especially when handling multiple concurrent workloads.
VDUSE: A technology that allows the creation of Virtio devices in user space, presenting them to virtual machines through the vDPA mechanism. This method enables block devices to operate with high performance and low latency, leveraging data path acceleration benefits. VDUSE also simplifies the development process by eliminating the need to modify or load modules into the Linux kernel.
While other methods like ublk exist, their limited support in virtual environments means that we do not focus on them in our solution. Instead, we prioritize the methods that provide the highest performance and scalability for cloud-based AI workloads.
Comparing the methods
To better understand the performance benefits of xiRAID Opus, we compared it to MDRAID using a set of RAID configurations in a controlled testing environment. The test environment included:
xiRAID Opus RAID 5 (23+1 drives) vs MDRAID RAID 0 (24 drives)
Single Virtual Machine: 32 VCPUs, 32GB of RAM
Operating System: Rocky Linux 9 with kernel-lt (6.10)
Test Tools: FIO v3.36
Workloads: 4k random reads, AIO (asynchronous I/O), direct I/O, full stripe writes
This configuration allowed us to evaluate the efficiency and performance of xiRAID Opus in random read and sequential write operations, comparing it directly with the MDRAID setup.
Passing shared block volume to 1 VM - Random read
In a workload of 1 job / 1 IO depth, xiRAID Opus (vhost-user-blk) achieves an impressive write of around 8 GBps, outperforming other solutions by almost 2x.
As the workload intensifies to 8 jobs / 32 IO depth, xiRAID Opus maintains strong performance, reaching approximately 70 GBps. While other solutions begin to close the gap under higher concurrency, xiRAID Opus still provides higher scalability and has the capacity to manage more complex storage tasks efficiently.
Passing shared block volume: Methods Comparison
SOCIAL SHARE CARD GENERATOR