Life in the Fast Lane

Optimisation for Batch and Interactive Jobs

Nikola Marković, Boemska Ltd.

About the Presenter

Nikola Marković is CTO at Boemska, a London-based SAS Silver Partner consultancy. He is the product architect for Enterprise Session Monitor for SAS, and and the Boemska HTML5 Data Adapter for SAS.

 

Nik specialises in the design and development of enterprise web applications built on the SAS BI platform, and dabbles in SAS architecture, performance optimisation and root cause analysis. He often enjoys writing about himself in third person.

SAS Jobs

SAS Jobs

or

Addressing the RAM

RAMFS

Unlimited. Similar to SASFILE. Overuse it and the OOM killer will kick in.

TMPFS

Successor to RAMFS. Can be hard limited, and can be resized on the fly. Uses VM, so will page to swap.

Kernel Module (brd)

Space is presented as fixed size block device. Needs formatting, and needs a filesystem. Safer but more space wasted.

Candidate Selection

Candidate Selection

Single Lane
Fast Lane

What is a Fast Lane Load Balancer?

Cost Based
Load Balancing

Load Balanced Fast Lane

Somebody ask about why we don't just increase memsize

memsize 150G

sortsize 150G

real time 14:43.14 
cpu time 24:44.02

disk read 103,073MB
disk written 103,089MB

100GB dataset

 

memsize 149G

sortsize 149G

real time 18:44.07
cpu time 36:34.22

disk read 103,073MB
disk written 209,947MB

100GB dataset

 

memsize 1G

sortsize 1G

real time 10:22.52
cpu time 19:11.47

disk read 103,073MB
disk written 209,947MB

100GB dataset

 

memsize 1G

sortsize 1G

real time 09:15.35
cpu time 18:09.22

disk read 103,073MB
disk written 103,094MB

100GB dataset

TMPFS UTILLOC

normal

biggest job 31m 07s
avg smaller job 25m

biggest job 23m 06s
avg smaller job 17m

fastlane

Runtime Comparisons

Server Performance (normal)

Server Performance (fastlane)

Disk Performance (normal)

Disk Performance (fastlane)

github.com/boemska/fastlane

(give me a few days to document it properly)

What's next?

codebase will be a community project, at

fastlane

By Nikola Markovic

fastlane

SAS Global Forum 2016 presentation

  • 733