Hmm.. I wonder if there are any potential performance gains by setting the page size to the MMC / SD block size (if possible) or forcing the kernel to only swap, say, 64 pages at a time. Is it possible the kernel is writing out 4kb pages in separate I/O operations?