rough estimate of kadmin addprinc performance?

[Chris Hecker]

> An enterprise-quality SSD (that can safely flush buffers to flash
> under power failure conditions) would probably net you a ~100x
> performance increase.

Yeah, this was my thought as well, although only in the randkey case.
I'll need to use more cores if the fsyncs go away in the pw case.

I'm going to upgrade the machine and see how it scales.  I'll post to
this thread with results (not that they're anything more than anecdotal,
but anyway).

Chris

On 2012/08/14 14:29, Nico Williams wrote:
> On Tue, Aug 14, 2012 at 2:58 PM, Chris Hecker <checker at d6.com> wrote:
>> I just tested creation with randkey for the sake of completeness (I need
>> the passwords), and yeah, it looks like this:
>>
>> CPU
>> 11%       slapd
>> 4%       kadmind
>>
>> but, slapd is 97% DSK, so I guess it's the fsync's.  It's basically the
>> same speed (10 ank/sec), which makes sense, since the previous test with
>> passwords wasn't 100% on the CPU, so it was disk bound as well on ldap.
>>
>> Deletes are also 10/sec.
> 
> Trace slapd to check how many fsync()s per-write it's doing.  I
> suspect it's doing more than a single fsync() per-write operation --
> assuming that you have a typical 7200rpm disk it must be doing quite a
> few fsync()s per-op!  You might research alternative slapd backends
> and/or using SSDs.  An enterprise-quality SSD (that can safely flush
> buffers to flash under power failure conditions) would probably net
> you a ~100x performance increase.
> 
> Nico
> --
>