> My worry about hardware RAID controllers is what if it breaks? Would > you > need to replace it with the same make, model, and firmware rev level to > recover your data?
I don't question it; I assume the answer is yes, if it breaks you must replace it with the same model. My approach to solving this is to always get the warranty on that product, so if it breaks I can always get the same model replacement from the mfgr. > Anyway, with multi-core processors, the CPU load from software raid is > not > very important. CPU load is not my concern either, in software raid. The three reasons I steer clear of software raid are: #1 If a disk goes bad, I expect hotswappable drives, with a red blinking light. #2 Not based on cpu strain, better performance. I believe every OS will write files, and in the software level, the kernel will not tell the application that a write has finished, until the write has finished. At least it keeps the written data in memory for subsequent reads, but the write is blocking. With the hardware raid controller, the hardware will tell the kernel it's written before it's written. (If you have write caching turned on, which it is not, by default with 3ware.) #2a It is easy to measure the performance difference with/without write caching enabled in the hardware. I've done this many times, and I encourage you to try it. The most dramatic thing is to "time mkfs" but the most realistic thing is to run iozone with something like 20G (this will take overnight; it doesn't just write 20G, it writes 128k, then 256k, then 384k, etc, write, rewrite, rewrite, etc etc) #3 Suppose you find a way to enable the above mentioned write caching in software kernel. Sometimes the kernel is then telling applications data's written that's not yet written. When a kernel crashes (rare but nonzero) there is more corruption than there would otherwise have been, if the cache was controlled by hardware which continues to write-out the contents of cache. I don't know realistically how important this feature is, but it has only one potential; it must be equal or better than software write caching. _______________________________________________ bblisa mailing list [email protected] http://www.bblisa.org/mailman/listinfo/bblisa
