-Kernels Must And Can Be Small

For a general acceptance, -kernels must be fast and not burden applications. For ful lling these conditions, cache architectures require -kernels to be small. The L4 -kernel shows that smallness can be achieved. 1. -kernels must be small This is not obvious. Most rst-generation -kernels were large; typically they need 300 Kbyte of code and 140 system calls. Some of their architects argued that ` ' in this context means `lower level' and not `small size'. Demanding smallness radically di ers from this approach. It could (and in fact it does) change -kernel technology dramatically. Why should -kernels be as small as possible? (We avoid the term \minimal" because of its mathematical implications.) The reasons are performance, exibility and perhaps correctness. 1.1. \A non-small -kernel is not fast." The most relevant performance costs of a -kernel result from its cache consumption. If a frequently invoked kernel operation accesses a substantial part of the primary cache (\ oods the cache"), the user is punished twice. First, the kernel operation itself is degraded by cache misses, since it must displace user code and data. Second, the user program has to pay for additional cache misses, since it must re-establish its cache working set. From Chen's [1993] measurements for a Mach benchmark on a DS 5000/200, we can calculate that on average 20% of the system cache misses are caused by user-kernel competition. We expect a substantially higher number for more up-to-date 30 Saw Mill River Road, Hawthorne, NY 10532, USA GMD SET{RS, 53754 Sankt Augustin, Germany hardware and faster (but non-small) kernels. On modern processors, the mentioned cache misses might consume up to 5 ( ve!) times as many cycles as the kernel code required for execution in the ideal case. What happens if the kernel working set is reduced, say from 75% to 15% of the cache size? The small kernel needs only 1/5 of the instructions. Since -kernel operations usually execute very few loops, we can expect a corresponding 4 to 5 times speed improvement. There is a good chance that competition between user working-set and kernel working-set is substantially reduced. Ideally, frequently invoking a -kernel operation does neither cost kernel-level nor subsequent user-level cache misses.