Carlo Codes MCNP-4A and 4B on dierent computers C R EDWARDS, BSc, MPhil and P J MOUNTFORD, PhD, FIPEM Medical Physics Division, North Staordshire Hospital (Royal Inrmary), Princes Road, Hartshill, Stoke-on-Trent, Staordshire ST4 7LN, UK Abstract. MCNP (Monte Carlo N-Particle) is a Monte Carlo transport code which has been of widespread use in modelling the dosimetry of ionizing radiations. The most recent version (4B) features improved electron transport compared with the previous version 4A. The processing time required by a number of computing systems to carry out X-ray and electron transport calculations using both versions of the code was compared. Version 4A was installed onto a Dec Alpha Server 8200, a personal computer (Pentium 90 MHz), and a Sun Sparc20, 10, 4 and 1+. MCNP-4B was also installed onto the Sun Sparc20. The benchmark tests consisted of determining the transmission of 2 MeV X-rays and 30 MeV electrons through lead. It was found that the Dec Alpha Server 8200 was the fastest computing platform, and the Sun Sparc1+ was the slowest for both tests. The dier- ence in computational speed between dierent platforms was not matched by the corresponding dierences in price. The time required by version 4B to complete the X-ray and electron benchmark tests was found to be 1.4 and 2.3 times greater than version 4A, respectively, without any dierence in the results of the calculation for each type of radiation. This suggests that in cases where computing time is important, it may be preferable to use version 4A instead of 4B. The Monte Carlo method is a computational technique based on statistical physics, which can be used to analyse the dosimetry of ionizing radia- tions [1]. An important factor in running any Monte Carlo code is the computing time necessary to trace a sucient number of particle or photon histories which will reduce the uncertainty to the required level [2]. This requirement could have a signicant inuence on computer purchasing deci- sions depending on whether access is required to a large platform, such as a network server system, or whether a less sophisticated system such as a work- station or a desktop personal computer (PC) is adequate. MCNP (Monte Carlo N-Particle) is a general purpose Monte Carlo code written in the Fortran and C languages [3]. Although MCNP-4A has been used for modelling X-ray transport and dosi- metry [4], its standard of electron transport was below that achieved by other Monte Carlo codes (e.g. EGS4) [5]. There were two limitations to this modelling in MCNP-4A. Firstly, the limited tabu- lation [6] in the code of the Landau energy loss dis- tribution results in accurate mean energy loss modelling only in the region of 1 MeV [7]. Secondly, the broadening [8] applied to this distri- bution becomes inaccurate at short path lengths [7]. Since July 1997, version 4B has been made available and contains an improvement to the modelling of electron energy loss. The range of electron path lengths has been expanded to provide accurate modelling over a greater range of ener- gies. Moreover, application of a correction factor [9] to the approximation used for the energy loss cross-section has led to improved modelling of electron energy loss, and the photon interaction cross-section library has been revised [7]. The cost-speed relation of the earlier version (MCNP-4A) has been reported in two studies [2, 10]. However, neither study used one of the more recent network server systems or Sun workstations and only version 4Awas tested. The purpose of this study was to compare the time taken by a network server system, a range of Sun workstations and a PC to calculate an X-ray and an electron transport benchmark test with MCNP-4A, and to compare the time required by version 4A and 4B to execute each test. These benchmark tests were chosen because they are based on experiments [11, 12] which have been cited in authoritative recommen- dations and data sources for radiation protection [13^15]. Materials and methods MCNP-4A was installed onto the following computing platforms whose operating systems Received 16 January 1998 and in revised form 14 July 1998, accepted 11 August 1998. The British Journal of Radiology, 72 (1999), 196^200 E 1999 The British Institute of Radiology 196 The British Journal of Radiology, February 1999 and processors are shown in Table 1: a Sun Sparc20, a Dec Alpha Server, a Pentium 90 PC, a Sun Sparc4, a Sun Sparc10, and a Sun Sparc1+. MCNP-4B was also installed onto the Sun Sparc20 computer. An X-ray benchmark test was designed to repre- sent the experiment performed by Karzmark and Capone [11] and modelled the history of 10 6 photons as they passed through various thick- nesses of lead (2^28 cm) (Figure 1a). To reduce computing time, the 6 MV polyenergetic X-ray beam used in their experiment was represented as a 2 MeV monoenergetic X-ray beam, since this lat- ter energy corresponded to the mean energy of the polyenergetic beam [16]. The beam was assumed to originate from a point source at a distance of 2.6 m from the lead block and had a divergence of 6.5. Air scatter between the X-ray source and the lead block was ignored with the transmission determined by calculating the number of photons passing through a circular eld of radius 2.5 cm. The electron benchmark test (Figure 1b) repro- duced the experiment performed by Loevinger et al [12]. Due to the longer time required to model electron transport, the history of only 10 5 electrons was modelled as they passed through various thicknesses of lead (0.01^0.1 cm). The 30 MeV electron beam generated during the original experiments was represented as a monoenergetic source, with the transmission then determined by calculating the number of electrons passing through a circular eld of radius 5 cm at a distance of 2 m from the scattering material. Results The results of the X-ray transmission calcula- tions generated by MCNP-4A and 4B on the Sun Sparc20 computer are shown in Figure 2, with the values of absorption coecient, half value depth and one-tenth value depth shown in Table 2. The X-ray calculation times and relative speeds of each computer normalized to that of the Sun Sparc20 operating version 4A are listed in Table 1. It can be seen that the Dec Alpha Server 8200 was the fastest computer platform with the Sun Sparc1+ being the slowest. When the X-ray calculations were repeated on the Sun Sparc20 using version 4B, the computing time was 39% greater than the time required by version 4A. The results of the electron transmission benchmark test generated by MCNP-4A and 4B on the Sun Sparc20 compu- ter are shown in Figure 3, with the calculation times and relative speed of versions 4A and 4B also shown in Table 1. It was found that the time required by version 4B for the electron calculation was greater than that required by 4A by a factor of 2.33. Discussion Hendricks and Brockho [10] used the 25-pro- blem test set included within the MCNP-4A Table 1. Time taken to perform the X-ray [11] and electron [12] transmission benchmark test, normalized to the com- puting time taken by version 4A to perform the calculation on a Sun Sparc20 Computer Operating Processor MCNP Benchmark System (MHz) X-ray Electron Time Relative Time Relative (h:min:s) speed (h:min:s) speed Sun Sparc20 Solaris 2.5 150 4A 0:13:16 1.00 8:24:50 1.00 Sun Sparc20 Solaris 2.5 150 4B 0:18:29 0.72 19:21:24 0.43 Dec Alpha Digital UNIX 23300 4A 0:04:09 3.20 Server 8200 Pentium 90 MS-DOS 6 90 4A 0:31:41 0.42 Sun Sparc4 Solaris 2.5 70 4A 0:39:10 0.34 Sun Sparc10 Solaris 2.5 50 4A 0:39:55 0.33 Sun Sparc1+ Solaris 1.5 25 4A 3:16:23 0.07 (a) (b) Figure 1. Schematic diagram of (a) the X-ray attenua- tion benchmark and (b) the electron attenuation bench- mark modelled by MCNP-4A and 4B, based on the experimental arrangement of (a) Karzmark and Capone [11] and (b) Loevinger et al [12]. Short communication: A Monte Carlo comparison 197 The British Journal of Radiology, February 1999 package to compare the processing speeds of a number of computing platforms, and found that a Sun Sparc2 workstation was approximately 13% slower than a 66 MHz Gateway-2000 PC (model 4DX2-66V). However, they did not test any of the computing platforms used in this study. A speed test by Al-Aan [2] using version 4A of MCNP showed that for modelling the transmission of 6 MV X-rays through a water medium, a PC with a Pentium 90 MHz processor was faster than a Sun Sparc4 by a factor of 3.0. However, it was found that for the X-ray benchmark used in this study (Table 1), this type of PC was faster than a Sun Sparc4 by a factor of 1.2. Possible causes of the dif- ference between these factors, and between the times taken to calculate each benchmark, may include the compiler and ags used during compi- lation [17] of the code, the construction of the pro- cessors and their mother boards, as well as dierences between the features of the two X-ray benchmark tests, such as the number of incident photons, and the attenuating material which will in turn aect the number of interactions modelled per incident photon. The Sun Sparc20 was faster than the Sun Sparc1+ and the PC (Pentium 90 MHz) for the X- ray benchmark test by a ratio of 14.8:1 and 2.4:1, respectively. However, this increase in speed was not matched by the relative cost of the two systems, since the current (late 1997) United Kingdom cost of the Sun Sparc20 is greater than that of the PCby an approximate ratio of 9:1. Comparison of the Dec Alpha Server 8200 and the Sun Sparc20 work- station shows that the former was faster by a ratio of 3.2:1. However, as before, this improvement in speed of the Dec Alpha system was not matched by the relative cost of the two systems, since the United Kingdom cost ratio (late 1997) of the Dec Alpha Server 8200 and the Sun Sparc20 is approxi- mately 21:1. There was good agreement between the values of absorption coecient, half value depth and one- tenth value depth measured by Karzmark and Capone [11] and the values calculated by both ver- sions of MCNP on the Sun Sparc20 (Table 2). The uncertainty associated with the calculated data was found to be 63%. As the penetration depth increased, the calculated results began to deviate below the published data (Figure 2 and Table 2) [11]. This was probably caused by the use of a 2 MeV monoenergetic X-ray beam in the benchmark test rather than the polyenergetic beam used by Karzmark and Capone [11]. The dierence between the times taken by ver- sion 4A and 4B to carry out a calculation will depend upon the same hardware, software and benchmark factors described above. For instance, when comparing the speed of versions 4A and 4B on a Pentium 200 PC using the Lahey F77-EM/32- 5.2 Fortran compiler [Al-Aan IAM, Private com- munication, 1998], Love et al found version 4B was only 5% slower than version 4A when calculating the percentage depth dose variation of 6 MV X- rays in a water phantom [18]. Whereas, version 4B was found to be 39% slower than 4A when cal- culating the transmission of a 2 MeV monoener- getic X-ray beam in lead on a Sun Sparc20 computer using the Sun Sparc Fortran 4.0 compi- ler (Table 1). Table 2 shows that for the X-ray Figure 2. Transmission of 2 MeV X-rays through lead (2^28 cm) calculated by MCNP-4A and 4B compared with experimental data produced by Karzmark and Capone [11]. Table 2. X-ray linear absorption coecient, half value and one tenth value depth derived by MCNP-4A and 4B for lead Karzmark MCNP-4A MCNP-4B and Capone [11] Absorption 0.41 0.4260.01 0.4360.01 coecient (cm ^1 ) Half value depth 1.7 1.6460.05 1.6360.05 (cm) One-tenth value 5.66 5.4460.16 5.4260.16 depth (cm) Figure 3. Transmission of 30 MeV electrons through lead (0.01^0.1 cm) calculated by MCNP-4A and 4B compared with experimental data produced by Loevinger et al [12]. C R Edwards and PJ Mountford 198 The British Journal of Radiology, February 1999 benchmark test used in this study, the improved electron transport modelling of MCNP-4B did not produce a signicantly dierent transmission when compared with the results generated by ver- sion 4A. However, this may not apply in all cases as Love et al have shown that for their dierent benchmark, version 4B produced results which were closer to measured data than version 4A [18]. It has been shown that the approximations for electron transport used by MCNP-4A may result in electron transmission being underestimated by up to 10% [7, 19], as compared with models based on the PRESTA scheme which can be accurate to within 1% [20]. However, because both versions of MCNP produced a deviation between the calcu- lated electron transmission and the measured data of Loevinger et al [12] (Figure 3), the cause may be due to the assumptions made in this study to reduce the overall time. For instance, Loevinger et al [12] used a lead applicator between the source and detector, and a lead shield to surround the detector, both of which would produce a number of secondary electrons. Neither the applicator nor the shield were modelled to reduce the overall com- puting time. It can also be seen from Figure 3 that even though the electron benchmark test took longer to calculate using version 4B due to its more detailed electron transport modelling (Table 1), there appeared to be no dierence between the transmission results produced by versions 4A and 4B. However, other studies measuring collision energy loss straggling of a 10 MeV electron beam have demonstrated signicant improvements to the transport modelling [21]. Conclusion Of the computers compared for X-ray transport calculations, it was found that the Dec Alpha sys- tem was the fastest, although the relative speed increase of this system compared to the Sun Sparc20 was not matched by its respective increase in cost. It was also shown that less costly systems might require prohibitively long computing times. However, the choice of computer system will depend upon individual judgement as to the accep- table values of cost and time. Improvements added to the electron transport modelling included in ver- sion 4B of the MCNP code have signicantly slowed the modelling for X-ray and electron beams compared with version 4Awithout necessarily pro- ducing more accurate results. Although version 4A produced acceptable results for the benchmark test used in this study, further studies such as those car- ried out by Love et al [18] are required to identify the range and type of radiation transport calcula- tions for which version 4B should be used. The authors would like to thank Mr Mark Engham (School of Physics and Space Research, Birmingham University) for his help and assistance with the X-ray timing benchmark. Advice is gratefully acknowledged from: Dr John Hendricks (Los Alamos National Laboratory, USA), Dr David Picton and Dr David Weaver (School of Physics and Space Research, Birmingham University), Dr Stuart Green (Regional Radiation Protection & Physics Service, Queen Elizabeth Hospital, Birmingham), Dr Derek Wanless (Computer Science Department, Keele University) and Mr Andrew Moloney and Dr Julia-Claire Handley (North Staordshire Hospital). Mr C R Edwards would also like to thank the West Midlands NHS Executive for their nancial support. References 1. Andreo P. Monte Carlo techniques in medical radiation physics. Phys Med Biol 1991;36:861^920. 2. Al-Aan IAM. A comparison of speeds of personal computers using an X-ray scattering Monte Carlo benchmark. Phys Med Biol 1996;41:309^13. 3. Briesmeister JF. MCNPa general Monte Carlo Code N-Particle Transport Code, version 4A. LA- 12625-M, Los Alamos, NM: LASL, 1993. 4. DeMarco JJ, Solberg TD, Wallace RE, Smathers JB. A verication of the Monte Carlo code MCNP for thick target bremsstrahlung calculations. Med Phys 1995;22:11^6. 5. Halbeib JA, Vandevender WH. TIGER, a one- dimensional multilayer electron/photon Monte Carlo transport code. Nucl Sci Eng 1975;57:94^5. 6. Borsch-Supan W. On the evaluation of the function w(l) for real values of l. J Res National Bureau of Standards 1961;65B:245. 7. Haleib JA, Kensek RP, Mehlhorn TA, Valdez GD, Berger MJ, Seltzer SM. ITS version 3.0: the integrated TIGER series of coupled electron/photon Monte Carlo transport codes. Albuquerque, NM, USA: Sandia National Labs, 1992. 8. Blunk O, Leisegang S. Zum energieverlust schneller elektronen in du nnen schichten. Z Physik 1950;128:500. 9. Seltzer S. Electron-photon Monte Carlo calcula- tions: The ETRAN code. Appl Radiat Isot 1991;42:917^41. 10. Hendricks JS, Brockho RC. Comparison of scienti- c computing platforms for MCNP4A Monte Carlo calculations. Nucl Sci Eng 1994;116:269^77. 11. Karzmark CJ, Capone T. Measurements of 6 MV X- rays 1. Primary radiation absorption in lead, steel and concrete. Br J Radiol 1968;41:33^9. 12. Loevinger R, Karzmark CJ, Weissbluth M. Radiation therapy with high energy electrons Part 1. Physical considerations 10 to 60 MeV. Radiology 1961;77:906^27. 13. National Council on Radiation Protection and Measurements. Radiation protection design guide- lines for 0.1^100 MeVparticle accelerator facilities, NCRP Report 51. Washington: NCRP, 1977. 14. International Commission on Radiological Protection. Protection against ionising radiation from external sources used in medicine, ICRP Publication 33. Oxford: Pergamon Press, 1982. Short communication: A Monte Carlo comparison 199 The British Journal of Radiology, February 1999 15. International Commission on Radiation Units and Measurements. Radiation dosimetry: electrons with initial energies between 1 and 50 MeV, ICRU Report 21. Washington: ICRU Publications, 1971. 16. Mohan R, Chui C, Lidofsky L. Energy and angular distributions of photons from medical linear accel- erators. Med Phys 1985;12:592^7. 17. Bielajew A, Rogers DWO. A standard timing bench- mark for EGS4 Monte Carlo calculations. Med Phys 1992;19:303^4. 18. Love P, Lewis DG, Al-Aan IAM, Smith CW. Comparison of EGS4 and MCNP Monte Carlo codes when calculating radiotherapy depth doses. Phys Med Biol 1998;43:1351^7. 19. Rogers DWO, Bielajew AF. Dierences in electron depth-dose curves calculated with EGS and ETRAN and improved energy-range relations. Med Phys 1986;13:687^94. 20. Nahum AE. Monte Carlo simulation of dosi- meter response. In: Proceedings of the 1997 Conference on Elocutions on Monte Carlo, in Monte Carlo; 20 September 1997. Monte Carlo, Monaco, 1997. 21. Hendricks JS. Introduction. In: Briesmeister JF, Hendricks JS, editors. Advanced MCNP. New Mexico: Los Alamos National Laboratory, 1997:1^19. Book review Imaging of Abdominal and Pelvic Anatomy. Ed. by F S Welli and M Manco-Johnson (Churchill Livingstone, Edinburgh) ISBN 04430512387 This is a Franco^American text which is well illu- strated with high quality radiological images and line drawings where appropriate. The book is divided into 14 parts, in total there are 36 chapters. Each part covers a single organ or system and each chapter details the normal radiological appearance of the particular system, using dier- ent methods of imaging, e.g. ``Part 8. The Pancreas'' contains three chapters, detailing sec- tional imaging, pancreatography and angiography. The images are of uniform high quality through- out with good labels and easy to follow legends. It also contains some embryological anatomy which is helpful for understanding subsequent anatomi- cal variants, for example embryology of the pan- creas. One of the aims of the book is to ``underscore the primacy of angiography and clas- sical radiolography in learning the anatomy of the abdomen and pelvis and to integrade results of the newer modalities into this knowledge base''. It therefore concentrates on angiography and con- ventional contrast studies, but also includes com- puted tomography, magnetic resonance imaging, transabdominal ultrasonography and a small amount of endoscopic ultrasonography. There is a wealth of information in this book. Most of it is available in other atlases but in this degree of detail it would probably be necessary to go to several dierent atlases and to this extent the book might be a useful addition to a departmental library. At a cost of almost 100 it is unlikely that a junior trainee would invest in this book in addition to the other atlases which are currently available for preparation for the rst part FRCR exam. J BARBER M B SHERIDAN C R Edwards and PJ Mountford 200 The British Journal of Radiology, February 1999