By alkagupta on Nov 19, 2008
"I respectfully disagree with your comment that there is no performance penalty with NFS...You show a 12% increase in Processing utilization which is an overall performance hit. You are doing approx. the same amount of work but chewing up more CPU...while your users are still the same, if scaled to 100% Util, NFS wouldn't allow as many users as LocalFS, because you are turning Blocks into packets into blocks, which is SLOW."As per the subject matter experts at Sun, this is a fairly common argument in regards to idle %. We don't know if it's wrong in this case or not - but idle is not always an indicator of future performance or perceived headroom, it is wrong to assume so. The reality is you don't know how much more you can get from your system as configured unless you push it to do so. % idle is a wrong metric to focus on (very common). The metric that matters in this case is supporting concurrent users with response times for all transactions falling within the guidelines. The NAS solution does this just fine for less $$. Also as per detailed data here, the io latency reported by iostat is the same for both DAS and NAS - therefore the statement that NFS is slow is not backed up by evidence.
To the question if NFS io ultimately cost more? Sure - but the missing point here is that it doesn't matter with this test. Might it hurt us down the road pushing to 100%? Maybe - but with a cheaper solution, and load scaled to 75% of the DAS solution, user response time was fine (this is what matters). And to use the analogy of idle, with over 25% cpu free, we believe it could scale another 800 users, equaling the DAS result. Once we arrange for hardware (better network and drivers) to push the load beyond 2400 users, we would know for sure. Pl. stay tuned.
Lastly, do most customer environments really run at 100%? Will the purported "NAS penalty" really affect them? This shows that even if they run at 75% (pretty high for most environments I've seen), it's not an issue. HTH.