On Fri, 19 Apr 2002, John Summerfield wrote: > Assuming 100 Mbit LAN, how does one decide whether one needs an > expensive server with high-performance SCSI? What are the guidelines? > > Given a basic contemporary Celeron with a recent ATA-100 disk drive > performing at (according to hdparm) 30 Mbytes/sec or better, it seems > certain to me that the LAN is slower than the drives. The result from hdparm is not a good indication for a file server as it comes from a _linear_ read. This was already mentioned in the thread about file systems, but I will repeat and extend it here: for a multiuser system (or a file server which acts on behalf of several users at the same time), several files are open at the same time and different places on the disk are accessed, so a good "seek time" is also important. Depending on the amount of intelligence for ordering these operations on the disk controller or in software (OS), the impact of "seek time" can vary. Another important factor is the amount of cache (both at OS level and on the controller) and the way it is used - that's why Western Digital introduced an 8 MB cache (instead of industry standard 2 MB) on some of its higher end IDE disks and typically hardware RAID controllers have some 16-128 MB cache. > For that matter, any hints for just how many clients such a basic > server can support? Yes, I know, "It depends." What do I need to know > and measure to be sure of putting in a box that "will do the job?" Tools like bonnie and iozone can give you a better view about what you can expect from your disk; they operate at a higher level than hdparm, so the results include the influence of the file system - usually that's what you want as the file system (and not the raw disk) is exported. The next step is to run them on the clients (or a Windows equivalent, don't ask me for references :-)) and watch the server. You can measure the network usage (f.e iptraf) and see if the network usage comes close to the maximum; if that's the case, then the network is probably the bottleneck and your disk subsystem is good enough. If not... -- Bogdan Costescu IWR - Interdisziplinaeres Zentrum fuer Wissenschaftliches Rechnen Universitaet Heidelberg, INF 368, D-69120 Heidelberg, GERMANY Telephone: +49 6221 54 8869, Telefax: +49 6221 54 8868 E-mail: Bogdan.Costescu@IWR.Uni-Heidelberg.De
