Exadata

[bgrenn]

I'm going to be following this up with more information on exadata.  I'm going to go back to presentations at openworld, and see what I can find that is public information.

To start with Kevin Closson is a GREAT resource, as I said at the meeting.. He is a top storage engineer, and prolific blogger.. I saw his presenation at openworld, and it was very good.. Here is his blog..
http://kevinclosson.wordpress.com/

FYI, there is a great discussion group on linkedin for exadata... Join it and learn about it !! They are comparing/contrasting the IBM offering with Netezza and exadata.

[bgrenn]

I've been putting a lot of thought into the exadata lately, and doing a LOT more research.

1st.  you have to to take a look at your database (the AWR report is good for this)...  are you I/O bound ?  Is your top wait event Disk ?  Then the Exadata will probably give you a big bang.  The more CPU you see in your database, and the more CPU spent hard parsing the less gain you will see..  

I think Exadata is definately a great answer for a datawarehouse (mostly big reads).. It basically the same thing as netezza and terradata offer.. Tried and true I/O centric box.

As far as OLTP, it may or may not give you that huge bang.  As I've said, the more CPU centric your application the less the gain.

Time will bear it out.. I will post more as I do more work on it.

[bgrenn]

I am working on building my own Exadata from commidity items on the HP website.. It appears that since Oracle bought SUN, their prices have been dissapearing.

So far I have

2 c7000 enclosures, These are 10u so I need 2 of them.. Each enclosure includes fans, 3 phase power, and 1 infiniband switches (40g).


2 X BL490C  nehalem blades 72G memory, extra network card, and infiniband card in the mezanines.


1 sb40c stocked with 6 disks for the nehalem blades

So far I have 2 x $60000 for the enclosures, 2 X 15,000 for the blades, and 1 x 5,000 for the disks.

3 HP p4300 with 12 Sas drives (1tb drives).. 30,000



TOTAL === 245,000

Price for a 1/4 rack exadata..  366,000

I guess the flash PCI memory is $40,000/storage server ??

[bgrenn]

New tact.. start with the v1 and build it with current technology.. According the specs (start with a 42u rack...  list price for a full RAC

8 Database servers   ==  $ 93,000
14 Storage servers   ==  $173,152
4 24 port switches   ==  $ 30,000
1 48 port eth. switch =  $  1,500

TOTAL hardware       ==  $300,000

Oracle list price   =    $650,000

[bgrenn]

I've been on exadata groups conversing with Others.. I found a hardware specialist in Norway who told me he priced the Sun hardware and found that a 1/2 rack is really only $300,000 US for parts.

No one is really arguing the Value of the Exadata.. The Speed numbers are so fast, the question is around the cost.  The 1/2 Rack Lists at 1.5 Million for $300,000 in parts.  That leaves 1.2 million for labor, installation and software.. And that assumes you already have a ULA for everything else (rac, compression, etc. etc.).

[bgrenn]

One of the items I did get a chance to test is AMD VS Intel for LIO's.. My test was to scan a table in memory 10,000 times, and average out the execution times.. Bottom line was that Nehalem did perform 35% faster than AMD shanghi for in memory lookups.. To me this is an important metric for the Exadata.. If you aren't currently running on nehalem, and you are running linux.. Beware that Exadata is faster because of the chip set.  Make sure you rule out any chipset differences when comparing so you get a better comparison.

[bgrenn]

Another intersting note...  Did you notice that with 11gr2 oracle once had flash cache memory for the SGA to support Exadata?? Did you notice it's not mentioned for the exadata anymore ?? They are using flashcache at the disk level instead. (It's write through cache by the way so it doesn't help writes )... The reason is that the Exadata software doesn't return blocks.. it returns just the row data asked for.. what does this mean ?? The SGA is useless anytime the Storage software kicks in.  Well mostly useless. 2 queries that return different columns, or result sets CAN'T share Data !!  Yes the SGA is out of the picture for these ...Oracle realized this, and abandoned the flashcache at the database level in favor of flashcache for the disk.  Think of the flash cache at the disk, the same way you thought of the SGA at the database level.
If you pay close attention, you will notice oracle talking a lot of "pinning" specific blocks in the flash cache, and allowing you to "pin" items in the flashcache... Starting to sound a lot like an SGA isn't it ??

Just saying something to think about....