Virtualisation and USP cache

I have USP's that virtualise a number of external arrays and each array in the USP has its own CLPR.  Recently I started noticing that one of the CLPR's was having a real hard time with one of the servers and was constantly destaging its cache.  It appeared that the other CLPR's were in some small way effected by this and I don't know why. 

HDS seem to recommend that you should have at least a CLPR for each virtualised system and it might be best practice to put it in a SLPR.  Just how much cache should be used for a virtualised system?  I am not sure and HDS seem to have different opinions on this too.  It could have something to do with the amount of cache on the virtualised system but we have been using 4 GB.  When I noticed the destaging on the CLPR, I doubled that and still had issues but it was better which led me to believe adding cache made a difference.  So with heavy usage, 15 TB of external storage might need 8 or even 12 GB of USP cache just to work well.

Currently the USP 1100 could have up to 256 GB of cache.  I don't know if the USP V can have any more but it sort of makes me wonder how you can virtualise 247 PB of storage.  I am about to virtualise a 40 TB array with the USP and I am eagerly waiting to see what figure HDS comes up with when we create a new CLPR.  That new system is going to work a lot heavier than the current virtualised array and it has three times the storage and only marginally more cache in its controllers.  I heard something like 0.5 GB cache for each TB but others say even less.  Considering the workload, I think I might use the 0.5 GB.  That's 20 GB for a new CLPR.  Just say I have an old XP1024 with 100 TB of disk.  Do I need 50 GB for a new CLPR? 

Somehow this seems to point to the fact that HDS may be dreaming of getting decent performance when they start virtualising large systems?  Assuming that 0.5 GB per TB, HDS must be pushing the boundaries with 1 PB of internal and external storage when heavy usage is envisaged.

Perhaps my figures are wrong but if what I am seeing with destaging in one CLPR causing issues with the other CLPR's, I would want to be absolutely sure that virtualisation is going to work on large arrays without impacting my heavily utilised USP.

I hope I am wrong...