I tend to agree with Greg here.  I've got  this huge raid(5) array
of IDE disks which are set to run at udma(1).  That means the underlying
disks should be doing transfers of about 10Mbytes/sec, I believe.  I'm
getting reads of 20-25Mbytes/sec through the array, which is a good deal
faster than the underlying individual disks.  I don't know what my write
performance is, because right now, the writes are generating so many crc
errors on one of the cables, that I'm sure things are skewed toward the
time it takes to write all of those printf statements to the klog socket.
-Brian
On Feb 26,  4:58pm, Greg Oster wrote:
} Subject: Re: About NetBSD server tuning! - Suggestions gratefully solicite
} sudog@sudog.com writes:
} > 
} > hello folks--so as my final bit I'd like to know if there are any
} > other suggestions for pulling up the performance on my
} > tens-of-millions-of-hits-a-day NetBSD servers..  Here's what I've got
} > so far:
} > 
} > 1. basic kernel tuning NKMEMPAGES, NMBCLUSTERS, etc based on avail ram
} > 2. put most-accessed stuff on non-RAID partitions
} 
} If there are major performance problems with RAIDframe, I want to hear 
} about them...  Putting data on even a RAID 1 set should result in writes 
} nearly as fast as the underlying media, and for many reads, faster reads
} than any of the underlying media could provide on their own...  
} (RAID 5 has it's own problems with "small writes", but that can sometimes 
} be overcome with appropriate RAID and filesystem tuning...  of course a lot
} of this depends on the filesystem workload, but I certainly don't agree with
} your claim that "most-accessed stuff should be on non-RAID partitions"
} in the general case...)
} 
} Later...
} 
} Greg Oster
} 
} 
>-- End of excerpt from Greg Oster