I'd like to second a couple of comments made recently:
    * If they don't regularly do so, I too encourage the ZFS, Solaris 
performance, and Sun Oracle support teams to sit down and talk about the 
utility of Direct I/O for databases.
    * I too suspect that absent Direct I/O (or some ringing endorsement 
from Oracle about how ZFS doesn't need Direct I/O), there will be a 
drain of customer escalations regarding the lack-- plus FUD and other 
sales inhibitors.

While I realize that Sun has not published a TPC-C result since 2001 and 
offers a different value proposition to customers, performance does 
matter and for some cases Direct I/O can contribute to that.

Historically, every TPC-C database benchmark run can be converted from 
being I/O bound to being CPU bound by adding enough disk spindles and 
enough main memory.  In that context, saving the CPU cycles (and cache 
misses) from a copy are important.

Another historical trend was that for performance, portability across 
different operating systems, and perhaps just because they could, 
databases tended to use as few OS capabilities as possible and to do 
their own resource management.  So for instance databases were often 
benchmarked using raw devices.  Customers on the other hand preferred 
the manageability of filesystems and tended to deploy there.  In that 
context, Direct I/O is an attempt to get the best of both worlds.

Finally, besides UFS Direct I/O on Solaris, other filesystems including 
VxFS also have various forms of Direct I/O-- either separate APIs or 
mount options for that bypass the cache on large writes, etc.  
Understanding those benefits, both real and advertised, helps understand 
the opportunities and shortfalls for ZFS.

It may be that this is not the most important thing for ZFS performance 
or capability right now-- measurement in  targeted configurations and 
workloads is the only way to tell-- but I'd be highly surprised if there 
isn't something (bypass cache on really large writes?) that can't be 
learned from experiences with Direct I/O.

Eric (Hamilton)

_______________________________________________
zfs-discuss mailing list
zfs-discuss@opensolaris.org
http://mail.opensolaris.org/mailman/listinfo/zfs-discuss

Reply via email to