A.P. Kalistratov – Post-graduate Student, Bauman Moscow State Technical University
G.I. Afanasyev – Ph. D. (Eng.), Associate Professor, Department «Information Processing and Control Systems», Bauman Moscow State Technical University
G.I. Revunkov – Ph. D. (Eng.), Associate Professor, Bauman Moscow State Technical University
P.S. Semkin – Associate Professor, Bauman Moscow State Technical University
One of the key characteristics of computing hardware is productivity, i.e., the ability to produce a certain amount of computing per unit of time. In the general case, under the performance evaluation it is supposed to obtain some numerical values that allow to rank the computing systems and (or) estimate the time costs for solving the problems. The performance of virtual machines depends not only on the resources allocated to it by the host system, but also on the use of these resources by other virtual machines running in parallel on this host system. The purpose of the work is to determine the most suitable scenario for using dedicated processor resources. The obtained results allow to draw a conclusion about the influence of parallel use of system resources on performance. To measure performance in different scenarios, the author made experiments with virtual machines under various settings of the host system hypervisor. During the work of the benchmark, the work of the trading company's information system is modeled. The measured parameter is the number of transactions performed for a specified time interval with the execution time less than a certain value. In general, you can see that the performance of a virtual machine depends not only on the number of threads allocated to the virtual CPU, but also on the access mode to the physical processor cache - in scenario 1, the cache is completely at the disposal of the virtual processor, in scenario 2 the cache is shared with the second kernel , and in scenario 3 - with three other threads. This explains the performance drop of 80% in scenario 3. In pursuit of the optimal allocation of resources, the authors also experimented with increasing the number of virtual machines while preserving resource allocation. Recall that the technology of hyper-threading allows you to divide each physical core of the processor into two streams of instructions. The maximum reduced performance is observed when resources are distributed between two VMs according to scenario 2. In this case, with the proportional resource allocation in scenario 3, the reduced performance is even lower than in scenario 1, implying the use of the same amount of resources, but without division between virtual machines. Thus, we can conclude that the performance of a virtual machine depends heavily on the parallel use of the processor. Depending on the distribution of system resources, the performance of a virtual machine can drop to five times. We can not exclude the influence of the hypervisor used, but this issue is not considered in this article. The paper examined the effect of the distribution of system resources on the performance of virtual machines. The numerical values of the performance drop were determined and the effectiveness of different distribution scenarios of limited system resources was compared. The obtained data allow to compare different formats of parallel use of the physical processor, i.e., fixing virtual processors on cores and threads and drawing conclusions about their applicability.
- Xisamutdinov R.A. Problemy' adekvatnosti oczenki proizvoditel'nosti vy'chislitel'ny'x sistem i kriterii razrabotki novogo pokoleniya testovy'x sistem // Vestnik UGATU. 2006. № 3. S. 142−147.
- Sravnenie proizvoditel'nosti virtual'ny'x mashin Hyper V 1 go i 2 go pokoleniya na primere raboty' 1S. URL = http://efsol.ru/articles/performance-comparison 1c.html (data obrashheniya 11.10.2017).
- Klemenkov P.A. Oczenka proizvoditel'nosti programmnogo obespecheniya v virtualizovannom okruzhenii na osnove atomarny'x testov // Trudy' Instituta sistemnogo programmirovaniya RAN. 2011. T. 21.
- Kant K. Introduction to Computer System Performance Evaluation. McGraw-Hill. NewYork. 1992.
- Hyper-Threading: «dva-v-odnom» ot Intel, ili Skry'ty'e vozmozhnosti Xeon // URL = http://itc.ua/articles/hyper-threading_dva-v-odnom_ot_intel_ili_skrytye_vozmozhnosti_xeon_9428 (data obrashheniya 11.10.2017).
- Maksimenko V.N., Filippov A.A. Analiz i sravnenie mikroarxitektur SMP i numa dlya sozdaniya vy'chislitel'ny'x sistem platformy' IMS // T Comm. 2010. № 10. S. 55−57.
- Fisun V.A. Obrabotka videodanny'x v sistemax speczial'nogo naznacheniya sredstvami vy'chislitel'noj texniki s ogranichenny'mi resursami // Specztexnika i svyaz'. 2011. № 3. S. 62−63.
- Pogue Charles, Anil Kumar, Douglas Tollefson, Steve Realmuto. Specjbb2013 1.0: an overview // Proceedings of the 5th ACM/SPEC international conference on Performance engineering. ACM. 2014. P. 231−232.
- Shipilev A., Keenan D. SPECjbb2012: updated metrics for a business benchmark // Proceedings of the 3rd ACM/SPEC International Conference on Performance Engineering. April 22−25 2012. Boston (Massachusetts. USA).
- Kalistratov A.P., Zaikin S.I., Kuzovlev V.I., Semkin P.S. Ispol'zovanie snapshotov pri organizaczii uchebnoj sredy' na baze virtual'ny'x mashin // Nauka i obrazovanie (nauchnoe izdanie MGTU im. N.E'. Baumana). 2016. № 7. S. 209−215. URL = http://old.technomag. edu.ru/doc/845284.html (data obrashheniya 11.10.2017).
- Marr Debbie, Frank Binns, D. Hill, Glenn Hinton, D. Koufaty. Hyper-threading technology in the netburst® microarchitecture // 14th Hot Chips. 2002.
- Chernen'kij V.M. Algoritmicheskij metod opisaniya diskretny'x proczessov funkczionirovaniya sistem // Informaczionno-izmeritel'ny'e i upravlyayushhie sistemy'. 2016. № 12. S. 11−21.