Introduction
Storage Area Networks (SANs) are essential for modern data centers, providing a flexible way to pool physical storage into logical volumes. Virtualization systems typically operate in one of two modes: in-band, where data flows through the virtualizer, often creating a performance bottleneck; or out-of-band, where the virtualizer is removed from the data path. While out-of-band systems offer better scalability, traditional implementations often suffer from high latency because host-side agents must query a central manager for every I/O operation. This paper presents an optimized out-of-band storage virtualization system for Solaris 10 that decentralizes the address mapping process to significantly improve performance and enable on-line volume expansion.
System Architecture
The proposed system is divided into two primary components: a central Virtual Storage Manager (VSM) and a host-side Solaris Agent. The VSM handles administrative tasks and metadata management, while the Agent is responsible for high-speed I/O processing.
As shown in Figure 2, the Agent architecture is split between user space and kernel space. The Comm Module (user space) communicates with the VSM to download and cache metadata. The Driver Module (kernel space) then uses this cached metadata to perform address mapping—translating logical block addresses to physical disk locations—entirely within the host's kernel. By moving the "intelligence" of address mapping to the edge, the system eliminates the need for per-I/O network calls to the VSM, effectively removing the central manager as a performance bottleneck.
Implementation and Volume Management
A key innovation of this system is its support for on-line volume expansion, allowing storage capacity to grow without service interruption. This is achieved through a coordinated multi-step process between the virtualization layer and the Solaris operating system. The system utilizes the lockfs command to temporarily suspend file system activity and growfs to expand the Unix File System (UFS) in increments. This staged approach ensures that the system remains responsive to applications even during large-scale resizing operations.
Furthermore, the Driver Module supports asynchronous mapping. It can decompose a single logical I/O request into multiple "child" requests that are sent to physical disks in parallel. This capability is particularly beneficial for striped volume configurations, where data is spread across multiple disks to increase throughput.
Performance Evaluation
The implementation was evaluated on a Solaris 10 testbed featuring a Sun UltraSPARC-III machine (Agent) and an Intel Xeon machine (VSM) connected via a Brocade Fibre Channel switch to an FC disk array. Benchmarking was performed using Iometer, configured with four worker threads and a mix of 80% read/20% write random access patterns.
The results, illustrated in Figure 5, highlight the significant performance advantages of the striped configuration. A striped volume consisting of four disks outperformed a linear volume by an average of 103.65%. The performance gap was most distinct with large I/O requests (2048KB and 4096KB), where the striped volume achieved more than three times the throughput of the linear volume. At its peak, the striped configuration provided a 215.01% performance increase over the linear setup.
Conclusion
By decentralizing address mapping and implementing efficient metadata caching within the host Agent, this out-of-band virtualization system significantly reduces I/O latency and improves overall scalability. The implementation on Solaris 10 demonstrates that high-performance storage management and on-line volume expansion can be achieved without the bottlenecks associated with centralized virtualizers. Future work will involve exploring metadata consistency across larger host clusters and porting the architecture to other operating systems and file systems.
Summary updated