ZFS raid write performance?

[Linda Kateley]

Is it possible that the suggestion for the "landing pad" could be 
recommending a smaller ssd pool? Then replicating back to a slower pool? 
I actually do that kind of architecture once in awhile, especially for 
uses like large cad drawings, where there is a tendency to work on one 
big file at a time... With lower costs and higher densities of ssd, this 
is a nice way to use them

On 6/23/15 8:32 AM, Bob Friesenhahn wrote:
> On Tue, 23 Jun 2015, kpneal at pobox.com wrote:
>>
>> When I was testing read speeds I tarred up a tree that was 700+GB in 
>> size
>> on a server with 64GB of memory.
>
> Tar (and cpio) are only single-threaded.  They open and read input 
> files one by one.  Zfs's read-ahead algorithm ramps up the amount of 
> read-ahead each time the program goes to read data and it is not 
> already in memory.  Due to this ramp-up, input file size has a 
> significant impact on the apparent read performance.  The ramp-up 
> occurs on a per-file basis.  Large files (still much smaller than RAM) 
> will produce a higher data rate than small files.  If read requests 
> are pending for several files at once (or several read requests for 
> different parts of the same file), then the observed data rate would 
> be higher.
>
> Tar/cpio read tests are often more impacted by disk latencies and zfs 
> read-ahead algorithms than the peak performance of the data path.  A 
> very large server with many disks may produce similar timings to a 
> very small server.
>
> Long ago I wrote a test script 
> (http://www.simplesystems.org/users/bfriesen/zfs-discuss/zfs-cache-test.ksh) 
> which was intended to expose a zfs bug existing at that time, but is 
> still a very useful test for zfs caching and read-ahead by testing 
> initial sequential read performance from a filesystem. This script was 
> written for Solaris and might need some small adaptation to be used 
> for FreeBSD.
>
> Extracting a tar file (particularly on a network client) is a very 
> interesting test of network server write performance.
>
> Bob

-- 
Linda Kateley
Kateley Company
Skype ID-kateleyco
http://kateleyco.com