Am 28.04.2014 16:39, schrieb Paolo Bonzini: > Il 28/04/2014 15:43, Peter Lieven ha scritto: >>> b) the allocationmap should be allocated even for out-of-range >>> opt_unmap_gran, using a granularity of 64 sectors in that case. >> >> Would the increase of the resolution bring any benefit? If we increase the >> resolution I think all sectors falling into the same >> physical cluster would end up with the same status. > > What if opt_unmap_gran is 32K or lower? In this case you're not using an > allocationmap.
As written I am fine with lowering this to 4K. Follow-Up or v3? Peter > >>> c) an opt_unmap_gran of 32K should be used by QEMU as a cluster size >> >> Do we have evidence of used cluster sizes on real hardware? I only know our >> Equallogic storages which have 30720 sectors. > > IDE does not specify the granularity, so an IDE disk will show a granularity > of 1 sector (but it has LBPRZ=0). > >>> d) the threshold should be dynamic and equal to iscsilun->cluster_sectors >> >> I don't think this is useful. For big cluster_sizes we will end up to never >> use the allocation cache. > > Right. > > Paolo
