Per-process I/O accounting in RHEL?

We're thinking about moving a series of high-volume 
mission-critical medical RDBMS from HP-UX to RHEL for cost 
savings and OS performance improvements, and have some 
RHEL/Linux questions...

Our beefy RDBMS servers are attached to a SAN and typically might 
have as many as 500+ long-lived server processes running 
concurrently, collectively processing 100 million+ queries/day.  
When we have an I/O bottleneck, it is very important that we are 
able to quickly identify which process(es) among many hundreds 
are most responsible and to what extent.  It is *not* enough to 
simply identify the disk subsystem getting the I/O or guess by 
who's using CPU, interrupts, or by looking at process states, 
nor is it enough to look at per-process IO stats after the 
process has completed.  We need to know the current quantitative 
I/O rates for each process while these processes are still 
running.

On HP-UX, HP's Glance application provides that per-process I/O 
accounting (giving us I/Os per second).  The relatively recently 
added CONFIG_TASK_IO_ACCOUNTING and CONFIG_TASKSTATS Linux 
kernel parameters provide what we need in Linux, but these do 
not appear to be enabled in RHEL 5.  I see per-process IO 
accounting appears to have been enabled by default now in 
Debian, Ubuntu, Suse, and others.

It's simple enough to rebuild the kernel.  But I understand that 
rebuilding our RHEL kernels voids our RHEL support.  There is 
considerable resistance to that in our organization.  I find it 
hard to believe it is 2008 and this issue was not solved long 
ago in Linux given how critical this function is to performance 
analysis and troubleshooting.  


So what's the shortest/best path to getting live per-process IO 
accounting in RHEL?  Can we pay Redhat to turn this on in a 
supported kernel for us?  Other thoughts/suggestions?
I'm aware of iotop, atop, and others, all dependent on these 
kernel parameters.  Are their other Linux tools that do this job 
without the kernel reconfig?

Thanks in advance.

Ed