Export Transformation Performance

The export transformation moves transformations outside of the kernel and into user space. However, crossing the kernel-user boundary incurs overheads that are not present in a basic kernel transformation network. Consequently, Section 5.2.1 measures the cost imposed by executing an identity transformation in user space, relative to a kernel identity transformation. Using an identity transformation avoids transformation-specific costs and isolates the overhead imposed by pushing a transformation into user space. This analysis is followed by Section 5.2.2, which examines the end-to-end user-space performance of three linear $\cal {O}$(N) transformations.

Section 5.1.1 shows that a write is slightly slower than a read for a file guarded by a kernel transformation. However, this difference is not noticeable during an access through the export transformation where the execution time of a read is virtually identical to that of a write. The small difference in times is masked by the large overhead imposed by crossing the kernel-user boundary. Therefore, in order to avoid redundancy, this section only presents performance results for transformations on reads.

Opening a transformation link that contains a user-space transformation takes roughly three times as long as an equivalent link that contains only kernel transformations. This penalty results from the additional time it takes to detect a non-kernel transformation, replace it with an export transformation, and find the shared library containing the transformation function. Specifically, opening a file guarded by a kernel identity transformation takes 38.8 $\mu$s, as shown in Figure 6. This is 3.4 times faster than the 132.4 $\mu$s required by opening a file guarded by a user-space identity transformation.