Your tone looks unnecessarily accusational. I gave brief information because starting an in-depth debate about the actual technical reasons didn't seem useful. What matters to the template text in partman-zfs is that ZFS has trouble with i386, not with 32-bit arches in general.
But if you insist I have no problem with elaborating. 2011/2/8 Philipp Kern <[email protected]>: > General-purpose registers, yeah, sure. This might create performance > problems, but i386 is as capable to get the processing done as amd64. Correct. > You might run into addressing trouble with virtual memory > management, due to the size of the caches ZFS wants. But then state > this instead of refering to unrelated issues at hand. Correct, although it's just part of the picture. > Could you elaborate on the MMU bit? I386 forces the kernel to share virtual address space with userland. If they were allowed to overlap, a cache flush would be required with every switch to/from kernel mode. The default in kFreeBSD for i386 is 1 GiB, which for some workloads is not enough and will lead to panic. The ZFS Tuning Guide in upstream wiki [1] recommends increasing this value to 2 GiB, but this can lead to trouble with some programs. TTBOMK this problem doesn't happen on other 32-bit architectures, where caches are aware of user/kernel separation, overlap may happen painlessly, and thus 4 GiB space is available (2 GiB are already enough for ZFS-enabled kernel). > I'm not aware of huge changes that > cannot be emulated by PAE, too. If that works it could be considered, but I heard PAE is very inefficient. This should be discussed in -bsd. [1] http://wiki.freebsd.org/ZFSTuningGuide#head-5572047de4f96e082e28f40c3855d7800b9033b2 -- Robert Millan -- To UNSUBSCRIBE, email to [email protected] with a subject of "unsubscribe". Trouble? Contact [email protected] Archive: http://lists.debian.org/AANLkTik-vnqYruKXsCH0RFfvomOYMFhGYUw�[email protected]
