Check out the new USENIX Web site. next up previous
Next: Cross-Subsystem Optimizations Up: IO-Lite Design Previous: Cache Replacement and Paging

Impact of Immutable I/O buffers

Consider the impact of IO-Lite's immutable I/O buffers on program operation. If a program wishes to modify a data object stored in a buffer aggregate, it must store the new values in a newly allocated buffer. There are three cases to consider.

First, if every word in the data object is modified, then the only additional cost (over in-place modification) is a buffer allocation. This case arises frequently in programs that perform operations such as compression and encryption. The absence of support for in-place modifications should not significantly affect the performance of such programs.

Second, if only a subset of the words in the object change values, then the naive approach of copying the entire object would result in partially redundant copying. This copying can be avoided by storing modified values into a new buffer, and logically combining (chaining) the unmodified and modified portions of the data object through the operations provided by the buffer aggregate.

The additional costs in this case (over in-place modification) are due to buffer allocations and chaining (during the modification of the aggregate), and subsequent increased indexing costs (during access of the aggregate) incurred by the non-contiguous storage layout. This case arises in network protocols (fragmentation/reassembly, header addition/removal), and many other programs that reformat/reblock I/O data units. The performance impact on these programs due to the lack of in-place modification is small as long as changes to data objects are reasonably localized.

The third case arises when the modifications of the data object are so widely scattered (leading to a highly fragmented buffer aggregate) that the costs of chaining and indexing exceed the cost of a redundant copy of the entire object into a new, contiguous buffer. This case arises in many scientific applications that read large matrices from input devices and access/modify the data in complex ways. For such applications, contiguous storage and in-place modification is a must. For this purpose, IO-Lite incorporates the mmap interface found in all modern UNIX systems. The mmap interface creates a contiguous memory mapping of an I/O object that can be modified in-place.

The use of mmap may require copying in the kernel. First, if the data object is not contiguous and not properly aligned (e.g. incoming network data) a copy operation is necessary due to hardware constraints. In practice, the copy operation is done lazily on a per-page basis. When the first access occurs to a page of a memory mapped file, and its data is not properly aligned, that page is copied.

Second, a copy is needed in the event of a store operation to a memory-mapped file, when the affected page is also referenced through an immutable IO-Lite buffer. (This case arises, for instance, when the file was previously read by some user process using an IOL_read operation). The modified page must be copied in order to maintain the snapshot semantics of the IOL_read operation. The copy is performed lazily, upon the first write access to a page.

next up previous
Next: Cross-Subsystem Optimizations Up: IO-Lite Design Previous: Cache Replacement and Paging
Peter Druschel