----- Original Message ----- > I'll see if I can track down why the rgrplvb option isn't performing as well. > I suspect the matter goes back to my first comment above. Namely, that the > slowdown goes back to the slowness of page cache lookup for the buffers of > the > rgrps we are using (not rejected ones).
Hi, I did verify that the vast majority of time is spent doing page cache lookups, in function gfs2_getbuf(). The time spent in gfs2_meta_read() and gfs2_meta_wait() are minimal--almost nil--presumably because the page is already in cache. The rgrplvb option isn't improving this because it still calls gfs2_rgrp_bh_get which still does all the page cache lookups for the rgrp we need. The rgrplvb option should, in theory, save us a lot of time when searching for a suitable rgrp, especially when there are multiple nodes doing allocations. I can't think of any reason why my patch would be incompatible with rgrplvb, but I'd like to ask Ben Marzinski to scrutinize my patch carefully and see if he can find any flaws in the design. Regards, Bob Peterson Red Hat File Systems
