CPU performance for Paramics simulation

 

Lianyu Chu

California Center for Innovative Transportation (CCIT), University of California, Berkeley

California ATMS Testbed, University of California, Irvine

11/29/2005

 

Introduction

 

This study aims to find the best CPU for Paramics simulation. The results may be valid for other simulation models such as Dynasmart and TransModeller and those programs that need intensive floating calculation.

 

Methodology

 

The network used for the test is I-880 network, which includes a 35-mile freeway and several adjacent arterials. It has about 6000 nodes, 143 zones, and four routing tables.  

 

There are two performance indices for the test:

(1) How long does it take to load the simulation network (i.e. loading time);

(2) How long does it take to simulate the network for 15 minutes (i.e. simulation time).

 

The Paramics version was version 5.1. The Exceed version was version 7.1. The test restricted to the computers / CPUs we have.

 

Results

 

Test 1:

 

The first test evaluated the performance of four CPUs, including Intel 2.8GHz and 3.6 GHz CPUs and AMD 64 3800+ and Dual Core 4400+.

 

 

CPU freq (GHz)

Cache1 (KB)

Cache2 (KB)

RAM (GB)

Loading time (sec)

15-min simulation time (sec)

Intel 2.8G    

2.8    

16

1024

1

50 

240

Intel 3.6G    

3.6    

16

1024

2    

55

185

AMD 64 3800+    

2.4    

64 

512

2

40 

158

AMD Dual Core 4400+    

2.2    

64    

1024

1

30

167


Results show that:

(1) Cache size determines how long it takes to load a simulation network. The cache1 size is the dominant factor. For the two CPUs with the same cache1 size, the CPU with bigger cache2 size performs better.

 

Comparing Intel and AMD cpus,. Computers with AMD CPUs load network at a faster speed than those with Intel CPUs. The reason is that Intel has smaller cache1 size and Intel’s bigger cache2 does not help much.

 

Comparing the two AMD CPUs with the same cache1 and different cache2, the CPU with bigger cache2 (i.e. Athlon 64 X2 4400+) size loads the network faster.

(2) CPU working frequency determines simulation speed.

 

Intel 3.6G computer simulates faster than the Intel 2.6G computer because the latter has slower CPU frequency. For the two AMD CPUs, the CPU frequency of Athlon 64 X2 4400+ is 2.2 GHz and the CPU frequency of Athlon 64 3800+ is 2.4GHz. As a result, the computer with Athlon 64 3800+ CPU simulates faster.

(3) Dual-core CPU excels in running several tasks

 

We conducted two simulation runs at the same time using both AMD 64 3800+ CPU and AMD Dual Core 4400+ CPU. For 15-minute simulation, it took 318 sec for the AMD 64 3800+ CPU computer to finish the two simulation runs while it took only 174 sec for AMD Dual Core 4400+ CPU computer. Since the simulation time for one simulation run is about 167 sec, Dual-core CPU seems to have two separate CPUs and each of them handles one simulation run.

 

Test 2:

 

The second test has two objectives:

(1)   Evaluate how RAM speed affects the loading time;

(2)   Compare the performance of Intel M CPUs for laptops and Intel CPUs for desktops.

 

CPU Freq. (GHz)

Brand

RAM

Memory Reading

loading time (sec)

10-min simulation time (sec)

Intel M 1.6G

1.6G

HP

512M

1650

72

181

Intel M 1.6G

1.6G

Fuji

512M

2119

58

177

Intel M 1.8G

1.8G

Dell Inspiration

512M

2182

45

157

Intel 4 2.8G

2.8G

Self-assembled

1G Dual

5248

50

175

Intel D 2.8G

2.8G

Sony

1G Dual

4628

63

182

 

Results show that

(1)   The computer with a faster RAM (regarding reading and writing time) loads network faster.

(2)   Regarding the simulation time, Intel M CPU at 1.6GHz and Intel 2.8GHz CPU has comparable performance.

(3)   Intel D CPU performs worse than Intel 4 CPU.

 


Conclusion

 
(1) AMD 64 CPUs are preferred for Paramics simulation.

 

Our test shows that AMD CPUs are better than Intel CPUs and dual-core AMD CPU excels in multiple simulation runs or multiple applications at the same time. The market prices (at 11/29/2005) for AMD CPUs are as shown in the table below. Although we did not have chance to test all available CPUs, a general conclusion we have is that higher CPU working frequency and higher cache size are needed for simulation.

 

The recommended CPUs are highlighted in the table, including AMD Athlon 64 4000+ and Athlon 64 X2 4400+. Although the two high-end CPUs, i.e. AMD Athlon 64 FX-57 and Athlon 64 X2 4800+, may perform best, they are too pricing and thus they were excluded from the recommendation list.

 

 

TargetDirect.com Price 11/29/2005 (US $)

AMD Athlon 64 3000+ (1.80 GHz) 512K

140

AMD Athlon 64 3200+ (2.00 GHz) 512K

170

AMD Athlon 64 3500+ (2.20 GHz) 512K

220

AMD Athlon 64 3800+ (2.40 GHz) 512K

300

AMD Athlon 64 4000+ (2.40 GHz) 1MB

350

Athlon 64 X2 3800+ (2x 2.00Ghz) 512K

350

Athlon 64 X2 4200+ (2x 2.20Ghz) 512K

430

Athlon 64 X2 4400+ (2x 2.20Ghz) 1MB

530

Athlon 64 X2 4600+ (2x 2.40Ghz) 512K

660

Athlon 64 X2 4800+ (2x 2.40Ghz) 1MB

820

AMD Athlon 64 FX-57 (2.80GHz) 1MB

1050

 

(2) For RAM, a higher rated memory should have better performance.

 

For example, if there are two types of RAMs, which are PC-4000 and PC-3200, the former is better. The recommended memory size is at least 1 GB in order to handle a big network. It is expected that a bigger size of RAM will help simulate a big network.