[ale] LAMP performance

[Pat Regan]

On Mon, 13 Sep 2010 15:11:24 -0400
Greg Freemyer <greg.freemyer at gmail.com> wrote:

> On Mon, Sep 13, 2010 at 2:51 PM, mike barnes <mdb3624 at gmail.com>
> wrote:
> > I have been asked to come up with a solution for a 6 sigma
> > application that can have up to 1 million transaction/hour.
> 
> 300 transactions a second.  Seems doable, but obviously high
> performance.

That's a tough question to answer because we don't know what is
actually involved in each transaction.  

> I wonder about SSDs.  They can only handle a fixed number of
> transactions per lifetime.  They use wear-leveling to smooth it out,
> but a 128KB erase block can typically only take 50,000 writes per
> lifetime.
> 
> Does anyone know if that ends up being a real world limitation or not?
> 
> At a minimum I'd want to research that aspect.  But I suspect new
> generation SSDs will handle it fine.
> 

Current generation SSDs should do fine.  I've read that Intel's X25 is
supposed to be able to do continuous writes for the period of the
warranty before it will fail.  I don't know if the math actually
supports that or not but the amount of real world writes you can put an
X25 through are definitely huge.

I think the biggest trick they use is the fact that 5-15% of the drive
isn't actually addressable (depending on make/model).

> > P --> Latest
> > Issues that I think I need to consider
> > Application level --> Data cacheing, bytecode cacheing
> > Database --> large memory foot print to hold most of database to
> > alleviate disk IO
> 
> SSD "reads" are not resource intensive and don't shorten the lifetime
> at all, aiui.  So I'm not sure you should pay both for a large memory
> foot print and SSDs.
> 
> ie. a fast SSD may be able to handle 25,000 read io's a second, so it
> should be able to keep up with 300 transactions/second even if most of
> the db is not ram resident.

The X25M in my laptop can pull off 16,000 seeks per second.  I have
never tested writes per second, though.

I do have Bonnie++ results in front of me from one of my older SATA
RAID 10 arrays (four 400 GB 7200 RPM disks).  They only managed 326.2
seeks per second.

> > What am I missing?

How large is the dataset?  Nothing says fast seeks like a dataset than
fits in RAM.  Memory is pretty cheap!  I know servers with 32 gig of
ram are quite cheap.  If your database fits in that you're golden :)

Pat