One reason that NAS is so successful is that data movement requirements are dropping as we move to a single-user, single-transaction model served over the web.
But, by the same token, web-based applications often produce scalability issues *far* beyond those with fixed endpoints.
I would be interested in the two of you commenting on two issues, derived from my general background in systems performance, not from specific knowledge of these technologies, but which seem like they might be relevant.
First, ethernet connections are not just limited by total bandwidth, but also by competition for that bandwidth. When there are a large number of users competing for the bandwidth, the resulting collisions and retries can lower the effective bandwidth quite dramatically over what the pipe could theoretically carry. SCSI, by contrast, is a much more managed protocol and can deliver sustained performance close to rated bandwidth when properly implemented. Is this contrast an issue in this debate?
Second, some storage architectures deliver excellent read performance, but are highly undesirable when a significant portion of the activity is writes. E.g., RAID 5 and a high transaction database are generally a very poor match since either write performance is poor or integrity compromised. Is there a difference in these architectures between read and write performance? |
