From Fri Aug 23 10:59:17 1996
Subject: NONMEM Performance Notes for Silicon Graphics Platforms

Dear NONMEM Users,

In the past there have been some queries on this forum about `NONMEM on SGI' systems and about `Best Platforms' for NONMEM. Thus, the following installation and performance notes should be of interest to current and prospective users of Silicon Graphics computer systems.

I will be also posting these on the NONMEM repository run by Steve Shafer and Jaap Mandema ( probably filed under NONMEM.DIR/SGI_TIPS.DIR) for future reference.

Tom Elken, Ph. D.
Advanced Systems Division
Silicon Graphics, Inc.

NONMEM Performance Notes for Silicon Graphics Computer Systems


Silicon Graphics POWER CHALLENGE 10000 computers (using MIPS R10000 microprocessors, MIPS is a subsidiary of Silicon Graphics Inc. (SGI)) provide excellent performance and accuracy on NONMEM jobs. These notes contain tips for obtaining optimal NONMEM performance on SGI platforms.

The R10000 has proved to be a well-balanced processor, providing very good performance on both integer and floating point codes. As others have noted in this forum, NONMEM relies more on integer performance than floating point performance. Like many programs which have evolved over the years, NONMEM has a considerable amount of branch logic (IF statements, etc.). The R10000 includes branch prediction, out-of-order and speculative execution features which allow it to perform well on existing codes with substantial amounts of program branching.

We have tested NONMEM on a number of customers' datasets, but the customers want these to remain proprietary. Thus, we cannot publish performance results. We would be very happy if a publicly available set of realistic (in size and complexity) benchmark problems were made available which computer vendors could use as measuring sticks for NONMEM performance.

These performance notes provide information in two areas that should result in considerably faster NONMEM runs for users of Silicon Graphics computers, while still providing accurate answers on a wide range of problems:

1) FORTRAN 77 compiler options (flags)
2) Buffer dimensioning recommendations


SGI/MIPS Processor type: f77 options
R10000/R8000: -O2 -n32 -mips4 -v6 \
-OPT:IEEE_a=3:ro=3:unroll_size=0:fast_sqrt -sopt,-o=0,-so=0,-ro=0
R5000/R4400: -O2 -mips2 -Olimit 1600 -align64 -sopt,-so=1,-r=2


Change the following statements in DAT1.f from

DIMENSION X(1000040)
DATA SIZE/1000040/


Explanatory details on installing/compiling NONMEM :

It is advisable to use double precision NONMEM.

No changes are needed to the default machine constants in BLKDAT because SGI uses the IEEE architecture for floating point numbers.



You can tell what microprocessor is in your system by typing `hinv' at a command prompt. You can tell what compiler version you have by inserting a `-version' option in any compiler command line.

MIPSPRO Version 6 or 7 f77 compilers:

f77 -O2 -n32 -mips4 -v6 -OPT:IEEE_a=3:ro=3:unroll_size=0:fast_sqrt \

Explanation of Flags:

-O2: There are four levels of optimization available with SGI's

MIPSPRO Compilers:

-O0 default - no optimization
-O1 local optimizations that can be done quickly
-O2, Turn on extensive optimization. The optimizations at this level are generally conservative, in the sense that they (1) are virtually always beneficial, (2) provide improvements commensurate to the compile time spent to achieve them, and (3) avoid changes which affect such things as floating-point accuracy.
-O3 Turn on aggressive optimization. The optimizations at this level are generally beneficial but occasionally hurt performance. With NONMEM, -O3 generally hurts performance.
-n32 The SGI compilers support three different Application Binary Interfaces controlled by the options -32 , -n32 , and -64. -n32 instructs the compiler to use N32, the new high performance 32-bit ABI which was introduced on Irix 6.2.
-mips4 Generate code using the full MIPS IV (e.g. R8000 or R10000) instruction set. This is the default on any R8000 based system. This implies -64 if -n32 has not been specified.
-v6 Specifies that the MIPSPRO Version 6 compilers be used. These are a no-cost option on Power Challenge R10000 computers. If you only have Version 6 compilers installed, you don't need this flag.
Specify the level of conformance to IEEE 754 floating-point arithmetic roundoff and overflow behavior. At level 3, allow arbitrary mathematically valid transformations.
Specify the level of acceptable departure from source language floating-point round-off and overflow semantics. At level 3, any mathematically valid transformation is enabled.
Specify the maximum size (instructions) of an unrolled loop (default 320). This indirectly controls which loops can be fully unrolled. When n=0, that means there is no maximum size for an unrolled loop.
Enable calculation of square roots using the identity sqrt(x) = x*rsqrt(x), where rsqrt is the reciprocal square root operation (-mips4 compilations only).
Turns off the KA preprocessor for the version 6 compilers. details of sopt options are found in the man pages for `fopt' and `copt', the standalone FORTRAN and C versions of this preprocessor.


MIPSPRO 5.3/6.2/7.0 f77 Compiler:

The R4400 and R5000 processors generally seem to operate well with the same set of compiler flags for any of these compiler versions:

f77 -O2 -mips2 -Olimit 1600 -align64 -sopt,-so=1,-r=2

Explanation of flags:

-Olimit n
Specify the maximum size, in basic blocks, of a routine that will be optimized by the global optimizer. If a routine has more than this number of basic blocks it will not be optimized and a message will be printed. An option specifying that the global optimizer is to be run (-O, -O2, -O3) must also be specified. n is assumed to be a decimal number. The default value for n is 1000 basic blocks.
Enforces alignment on 64-bit boundaries for 64-bit objects; all smaller objects must still be aligned on the corresponding boundaries.
-sopt is a special source-to-source optimization pass for FORTRAN programs. Options for sopt must appear in a comma separated list following the -sopt option without any intervening blanks.
Short name: -so=integer
Default value: -so=3
Controls the level of scalar optimizations that fopt attempts. The settings of interest are -so=0 (no scalar optimization), -so=2 (full scalar optimization), and -so=3 (full scalar optimization plus memory management transformations).
Short name: -r=integer
This option specifies the amount of round off error that the user is willing to tolerate as a result of optimization.
-r=0 means that no such variation will be tolerated.
-r=2 means that variation due to roundoff is acceptable, and so roundoff changing optimizations will be done. For example, the floating point expression A/B/C will be computed as A/(B*C).

For heavily floating point codes, SGI documentation suggests "-O3 -n32 -mips4" as a good base set of options for Irix 6.2 or newer versions of the operating system. It turns out that integer performance is more important on NONMEM (for most datasets) than floating point performance. Thus, we recommend -O3 or -O2 along with -mips2 as a base set of flags for R5000 and R4X00 microprocessors. -mips2 implies the -32 ABI as a default, so we haven't listed that flag in these instructions.


The greatest performance improvements to NONMEM can come from changing one number in one of the NONMEM base subroutines. It almost always will help for medium to large datasets to make the following change to the DAT1.f source file (or to the DAT1 subroutine if you haven't split the code into one file per subroutine):

Change the following statements from

DIMENSION X(1000040)
DATA SIZE/1000040/

The NONMEM executable will use about 8 MB more of RAM when it runs, but this is not a problem for most SGI workstations or servers. This change can often provide a 50% to 100% performance increase, and rarely, if ever, would there be any degradation.

This dimension will handle problems with sizes up to 50000 data records. If you have a larger problem, set this dimension to a 20 times m+2, where m is the number of records in your data file. (You will probably forget about this change later, so it is good to anticipate the largest data file you will typically use.)

Keywords: Platform, Computer, Compiler

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~
Thomas Elken
Supercomputing Applications Phone: (415) 933-6495
Silicon Graphics Computer Systems FAX: (415) 933-3562
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~