On Tue, 11 Aug 2009, Dave Miner wrote: > Alok Aggarwal wrote: >> Hi Peter, >> >> On Tue, 11 Aug 2009, Peter Tribble wrote: >> >>> My initial impressions of AI performance were that it was significantly >>> slower than jumpstart. Given that my major criticism of jumpstart is >>> how slow it is, that's a problem. >>> >>> (And I know how to speed jumpstart up quite a bit - simply use less >>> computationally intensive compression schemes for the packages. >>> Really, to get it up to scratch would need 2-3X improvement in what >>> Jumpstart can currently give.) >>> >>> So an actual test: >>> >>> Test client is a SunBlade 1500 with 1G RAM. >>> >>> AI, with the repo image lofi mounted: 71 minutes to install packages. >>> SXCE Jumpstart, off my usual (busy) jumpstart server: 44 minutes to >>> install packages >>> AI, repo copied to disk: 70 minutes to install packages. >> >> Just curious why you went with a repo on disk as opposed >> to a networked repo? The repo on disk is particularly not >> suited to streaming reads. >> > > The lofi-mounted vs. copied to disk numbers above would seem to indicate no > substantial impact there; it's not actually on DVD media, which is the real > read latency killer.
Ah, I overlooked the fact that's it's a hard disk and not DVD media. > I'll note that the AI times Peter quotes are similar to those found in the > internal testing of x86 virtual box instances done over http. The > corresponding Jumpstart numbers are very different, but that may be a virtual > box artifact. The jumpstart numbers we have internally are indeed very different (for the worse). Alok
