i n t r o d u c t i o n
   g o a l s
   m e t h o d s
   s t a t u s
   r e s u l t s
   p u b l i c a t i o n s
   c o o p e r a t i o n s
   c o n t a c t    u s
   l i n k s

Status

• Installation and performance testing of the Portable University Model of the Atmosphere (PUMA-2) on a local Linux-Cluster

• Installation of Climate Community System Model (CCSM-Paleo, version 1.4) on an IBM Power-4 which contains 8 frames with 32 CPUs) in collaborations with specialists of the Swiss Centre of Scientific Computing (SCSC) in Manno.

• Performance testing of the CCSM on the IBM.

• Comparison of PUMA-2 and CCSM-Paleo 1.4.

• Installation of CCSM on a local Linux-Cluster (max. 32 CPUs) and performance testing.

• Building up a collection of diagnostic tools for oceanic and atmospheric analyses (internal use only!).

Performance of PUMA-2

Here, the influence of the number of CPUs, different computer architectures, and compilers are used to test model performance.

Computer	Compiler	CPUs	avg. time per year	Addionals
Compaq Workstation	Compaq F90, version 5.3
XP1000 (500MHz)		1	27min
		2	33min	100Mbit network
Alphaserver 2100 (666Mhz)		1	24min
		2	16min	Dual-board
Linux PC (AMD1400)	pgf90	1	23min
Linux PC (AMD1800)	pgf90	1	19min
		2	12min	Dual-board
Linux PC (AMD1800)	Absoft F90	1	25min
		2	15min	Dual-board

Performance of CCSM-Paleo 1.4

Parameter used for testing are: total number of CPUs, optimization compiler options, number of CPUs per model component, and degree of parallelization (threading).

Tests with 8 CPUs: (cpl atm ocn ice lnd) = ( 1 4 1 1 1)

test name	threads	compile option	avg. time per month
opt2.old	1	-O2,-qstrict	37 min.
opt3.old	1	-O3,-qstrict	42 min.
opt3.2.old	1	-O3	41 min.
opt2	4	-O2,-qstrict	12 min. (NCAR standard setup)
opt3	4	-O3,-qstrict	15 min.
opt3.2	4	-O3	15 min.

- "threading" is the most effective factor.
- optimization shows neglectible differences.


Tests with 16 CPUs: -O2,-qstrict

test name	threads	cpl	atm	ocn	ice	lnd	avg. time per month
cpu16.1	4	1	8	4	1	2	14 min.
cpu16.2	8	1	8	4	1	2	9 min.
cpu16.9	12	1	8	4	1	2	stops after 9 months
cpu16.10	16	1	8	4	1	2	stops after 8 months
cpu16.3	4	1	8	4	2	1	13 min.
cpu16.12	8	1	8	4	2	1	stops after 7 months
cpu16.4	4	1	8	5	1	1	13 min.
cpu16.5	4	1	12	1	1	1	12 min.
cpu16.11	12	1	12	1	1	1	12 min. (but ranging 7-17 min.)
cpu16.6	4	1	10	3	1	1	atm = 10 CPUs never work
cpu16.7	4	1	6	7	1	1	16 min.
cpu16.8	4	1	4	9	1	1	16 min.
cpu16.13	8	2	8	2	2	2	9 min.
mrenold	6	1	6	6	2	1	16 min.
cpu16.14	6	1	6	6	1	2	15 min.

Comparison of PUMA-2 and CCSM-Paleo 1.4

CCSM-Paleo	PUMA-2
Advantages:
comprehensive model	low computational costs
coupled without flux-corrections
Disadvantages:
computationally expensive, but ~10 years per day is attainable	not yet coupled,   presumably with flux-correction
	poor climatology (esp. precipition, NAO centers are displaced eastwards)