POWER7/PowerVM vs x86/VMware/OracleVM for oracle DB hosting

A Colleague ( http://twitter.com/ZoranGagic ) sent me this great breakdown of x86 vs Power7 for hosting oracle...nice to have it all in one spot (below) also this http://virtualgeek.typepad.com/virtual_geek/2011/07/even-more-reasons-to-run-oracle-on-vmware.html makes a similar case and now that VMWare have softened up on the vSphere 5 memory tax there is not as much need to go 4.1 as per below

	HP c7000 with 16 x BL460 G8	IBM 770
CPU cores	16 x 16 = 256	26
Memory	16 x 144GB = 2304	512GB
Performance, SPECintrate2006	16 x 662 = 10,592	942
Price	est $200K	$452K (approx)
COD	None	18 cores/512GB, cost: $10K for 1 core/16GB
Price difference		2.3 x more expensive
Performance difference	11.2 x faster
Price /performance difference	25.7 x better price performance

This does not include virtualisation or OS costs, but numbers are overwhelming….

General purpose CPU - Intel Xeon E5-2600:

http://www.theregister.co.uk/2012/03/06/intel_xeon_2600_server_chip_launch/print.html

This will be in every major rackmount, blade server, standard chipset support 24 DIMM sockets (24 x 32GB = 768GB), 10 GbE etc….

(Sub $10K for 16 core server with over 144GB memory)

Xeon E5 (2.3b transistors, more CPU cores (8), less heat/power due to 22nm process, bigger L3 cache (20MB)) :

http://www.spec.org/cpu2006/results/res2012q1/cpu2006-20120211-19458.html

16 cores, 2 chips, 8 cores/chip, 2 threads/core

SPECint_rate_base2006 = 662 (41.3/c)

Power 770 with Power 7:

http://www.spec.org/cpu2006/results/res2010q1/cpu2006-20100208-09583.html

48 cores, 8 chips, 6 cores/chip, 4 threads/core

SPECint_rate_base2006 =1740 (36.2/c)

Xeon 5670 in HP BL460:

http://www.spec.org/cpu2006/results/res2010q2/cpu2006-20100524-11203.html

12 cores, 2 chips, 6 cores/chip, 2 threads/core

SPECint_rate_base2006 = 318 (26.5/c)

http://www.oracle.com/us/corporate/pricing/technology-price-list-070617.pdf

http://www.oracle.com/us/corporate/contracts/processor-core-factor-table-070634.pdf

Enterprise Edition Per-core licensing

Multi-core processors are priced as (number of cores)*(multi-core factor) processors, where the multi-core factor is:

IBM Power 7 CPU Pool with 8 cores = 8 x $47,500 = $380,000

Intel E5 with 16 cores = 16 x $23,750 = $380,000

Better to buy a low cost Intel E5 based server with maximum memory available (768GB) and choose VMware ESX 4.1 (not huge memory costs with ESX 5) and RHEL 6.2 on a farm of blade servers such as:

http://www.dell.com/us/enterprise/p/poweredge-m620/pd

HP BL460 G8 is available with Intel E5 CPU:

                                                                                                

 http://store.vmware.com/store/vmware/en_US/pd/productID.126818400?resid=T1iZSAoHAtcAACq6rR4AAAAG&rests=1331206471307

(VMware ESX 4.1 Enterprise Plus - Unlimited memory

VMware vSphere 4.1 Enterprise Plus for 1 processor (Max 12 cores per processor) + Production (24x7 for Severity 1 issues) 3 Year Support

$5,723.70

                https://www.redhat.com/apps/store/server/

                                                                2-sockets with unlimited virtual guests

                                                                Standard Subscription (1 year) $1,999

                                                                Premium Subscription (1 year) $3,249

Oracle Prod on physical database farm with RAC (or one node RAC to keep it very simple)….many instance on one OS image:

-        Large memory (Intel memory is cheap), separate SGAs for different instance – no memory contention

-        Separate LUNs so IO is well separated between instances, if there are ever any “problem instances” can easily move to another server

-        Huge CPU as hardware and Oracle costs are significantly cheaper

Oracle on VMware (other better supported solution is Oracle VM):

http://www.embarcadero.com/images/dm/technical-papers/technote-oracle-virtualization.pdf

The Internet map

The Internet map

An interactive map of teh interwebs!

3 Networky AIX 6/7 gotchas....

A colleague came across this tricky gotcha in AIX7 - his scp sessions were stalling on AIX 7.1 (VIO client) LPARs.

There is not much on the net about this one yet

By disabling tcp timestamp randomisation feature on both source and target VIO clients the randomised timestamp value is not used for setting the retransmission timer

Fix

 #no -o tcp_rand_timestamp=1

Here is the APAR http://www-01.ibm.com/support/docview.wss?uid=isg1IV13121 IV13121: TCP RETRANSMIT PROCESSING IS VERY SLOW.

One possible cause of an SEA on VIOS flip flopping from primary to backup is processor unfolding delays http://www-01.ibm.com/support/docview.wss?uid=isg3T1012941

Fix - stop folding
# schedo -p -o vpm_fold_policy=4

There is a fair bit on the net about this one...

Another one I'll put here (AIX 6/7 and Oracle 11g network stalling/delays)
With Oracle 11g came IPv6 support and even though a hostname may be resolved in IPv4 land a second lookup into /etc/hosts is done for the IPv6 address/hostname.

If the IPv6 address is not matched in /etc/hosts then it goes out to the DNS servers for IPv6 name resolution because of the (default) name resolution order in net service (/etc/netsvc.conf) was local then bind
Fix
Force the IPv4 ie in /etc/netsvc.conf change local, bind to local4, bind4

There is a LOT on the net about this one....

Graphical representation of machine generated data

In this plot Im running a (ruby) gltail of apache access log while I rip the site with a recursive wget Often with machine generated data (eg a syslog stream) there is too much to read every item and its coming in real time from multiple sources so you cant see the forest for the trees. By graphically showing the data the eye can see if something looks strange or different than normal. The human eye has a data bandwidth of ~10 Mbps and couple that with the brain and you have a kick ass data miner!

Ask For Forgiveness Programming

I recon THIS is how the brain mitigates Amdahl's law and why our brains are 'slow and inaccurate' when compared to a computer at adding numbers.

ACID/locking/mutex/semaphore etc just wont scale to billions of cores (neurons)

However a computer is 'slow and inaccurate' at identifying real word objects or doing stand-up comedy.

Using a GPU to hack(brute-force/dictionary)at a salted MD5 hash....

Ive been using john(JtR)for years and one week while my laptop fan was overblowing I thought: there must be a better way than cooking my CPU for a week,that GPU must be able to help so I jumped on the internets and found a few GPU based tools.

I tried hashgpu http://www.golubev.com/hashgpu.htm and hashcat http://hashcat.net/oclhashcat-plus/

I used the OpenCL version of Hashcat+ (there is a cdua version for nivdia cards but I have a ATI) to pick away at a salted MD5 hash.
A good wordlist is your best bet,even try and cewl candidate data for a wordlist before you resort to a bruteforce then you can run mp(on hashcat site too) to generate a brute force char stream to stdout

$mp64 -1 ?l?u?d ?1?1?1?1?1?1?1?1 #Lower,Upper,Digits for 8 chars

#pipe to hashcat (8 char Lower,Upper,Digits bruteforce...take a while)
$mp64 -1 ?l?u?d?s ?1?1?1?1?1?1?1?1 | oclHashcat-plus64 -m 500 hash.txt #dropped the rules mode 500 is md5 unix)

I ran a dictionary based attack on my hash with a ruleset that comes with Hashcat, then I use mp to generate 8 char passwords...and go to bed while it chews away on the hash....how good is this software...and its free

oclHashcat-plus64.exe -m 500 hash.txt example.dict -r rules/best.rule

here is my demo/screen shots....
http://www.youtube.com/watch?v=5LTs_mmne0Q