Performance test of new real time plasma control system (PCS) server environment based on MRG-R Yeonjung Kim a , S. H. Hahn a , W. R. Lee a , H.Han a , M. H. Woo a , B. G. Penaflor b and Y. S. Bae a a National Fusion Research Institute(NFRI), Daejeon, Korea b General Atomics, San Diego, CA, USA 1
Table of Contents 1 . Introduction - Why Performance T est ? - MRG-R 2. Method for MRG-R System Tuning 3. Conditions for Performance T est - Input values & setting - Measurement of Time Register 4. Results of Performance T est 5. Summary and Future Plan 2
Introduction(1) Why the performance test ? The real-time control environment of KSTAR Plasma Control System(PCS) requires fast control cycles up to 20 KHz, which is very difficult to support at the standard Linux OS scheduler. KSTAR uses a customized Linux kernel which is not influenced by any external interrupts in the real-time operation mode( rtmode) . The MRG-R[http://www.redhat.com/products/mrg/realtime] is an ITER-compatible real-time Operation System (OS) which showed the most excellent performance. We want to confirm that the MRG-R is suitable as the real- time OS environment for the KSTAR PCS. 3
Introduction(2) Red Hat Enterprise MRG Realtime (MRG-R) BENEFITS & FEATURES of MRG-R Highly deterministic performance High-resolution actions Full application compatibility with standard Red Hat Enterprise Linux. Uninterrupted services for high-priority processes. MRG-R replaces the standard Red Hat Enterprise Linux kernel in order to provide extra support for real-time computing. 4
MRG-R System Tuning (1) PCSRT6 Version : RHEL 6.3 (2.6.32-279.el6.x86_64) MRG-R (3.8.13-rt27.33.el6rt.x86_64) Disable the graphic desktop: Set runlevel 3 Disable the irqbalance deamon - Use chkconfig to ensure that irqbalance does not restart on boot. - Open /etc/sysconfig/irqbalance in editor and find the section of the file titled FOLLOW_ISOLCPUS and remove the # character. - change the values to yes . 5
MRG-R System Tunning(2) Booting by the MRG Realtime kernel (a) Open the /etc/grub.conf file and add crashkernel =128M (b) Add the parameter for TSC timer synchronizaton clocksource=tsc powernow-k8.tscsysn=1 (c) isolcpus = 13-23 is boot kernel option for CPU isolation. 6
Conditions for Performance T est(1) Run the test program (testrtmode.c) at the isolated CPU to match the conditions of KSTAR RT process running Input values setting - CPU number =13 - Test period for test program (NUMSECOND)=5 seconds Main settings - Number of sample cycles : 5000 or 20000cycle - Threshold value (=unit for control cycle) : 1/10/20/50 µs Method of performance measurement - Measures total running time more than the threshold value. - Do the test for threshold 10/20/50µs as well. 7
Conditions for Performance T est(2) Measurement of total running time by the Time Register Use the internal 64-bit clock counter as the time register. Read into Register(EDX:EAX) using the instruction of RDTSC (read time stamp counter)[http://en.wikipedia.org/wiki/Time_Stamp_Counter] Using the readclock() function call , get the count values. Counters are converted as physical time using the CPU clock cycle. (For pcsrt6, the CPU clock cycle is measured as 2300MHz) Calculation of the total running time total count values is divided by CLOCKMICROSECOND(2300) 8
Conditions for Performance T est(3) * Optimization of test program (testrtmode.c) while (NUMSECOND & max_sample_cnt< MAX_SAMPLE ) { Readtick(t_diff= t_current - t_last); /* get the count values of each section */ total_clock_val=+ t_diff; /*accumulation of running count values */ max_sample_cnt++; If(t_diff > THRESHOLD*CLOCKMICROSECOND) /* excess threshold*/ { save_elapsed_seconds[num_delays] = elapsed_seconds save_elapsed_milliseconds[num_delays] = elapsed_milliseconds save_t_diffs[num_delays] =t_diff/CLOCKMICROSECONDS; save_excess_time += save_t_diffs[num_delays]; num_delays++; /* number of excess threshold values */ } else num_lower_delays++; /* number of less than threshold values */ } 9
Conditions for Performance T est(4) Explain the Variable used the test program(testrtmode.c) Variable & synonym explain CLOCKMICROSECOND CPU clock cycle per 1µs =2300 max_sample_cnt num_delays + num_lower_delay MAX_SAMPLE(synonym) sample cycle(synonym) total_running_time Total_clock_val /CLOCKMICROSECOND Response time (synonym) average_excess_cnt ( num_delays/max_sample_cnt )*100 average excess rate(synonym) 10
Results of Performance T est(1) Result of test Basic test: only measure the counters Dummy load test : add the additional load using a printf() between two measurements in repetitive statement while(). Expected Total Running Time Threshold ( µs) sample 5000 cycle Sample 20000cycle 1µs 5000 µs 20000 µs 10µs 50000 µs 200000 µs 20µs 100000 µs 400000 µs 50µs 250000 µs 10000000 µs 11
Results of Performance T est(2) Average excess rate of basic test Threshold ( µs) sample 5000 cycle Sample 20000cycle 1µs 0 % 0 ~0.01% 10µs 0 % 0 % 20µs 0 % 0 % 50µs 0 % 0 % * Total test number is 10 times * No sample cycles behind the threshold are found at the first 5000 sample cycles 0.01% of sample cycles behind the threshold are found at the first 20000 sample cycles at threshold 1µs (20% of total test number) From the basic test, the MRG-R environment is capable to perform a single process up to 10µs(=100 kHz) of the control cycles without delay 12
Results of Performance T est(2) Average excess rate of dummy test Threshold ( µs) sample 5000 cycle Sample 20000cycle 1µs 0~56 % 0~89% 10µs 0 % 0 % 20µs 0 % 0 % 50µs 0 % 0 % * Number of total test is 10 times * Minimum value is 0, Maximum value is 89. Addition of the printf() (calls interrupts for display) makes the test program performance slower However, the test program did not have performance delay up to 10 µs of control cycle by the additional dummy load The 1 µs threshold is seriously affected by this interrupt calling 13
Summary and Future Plan Performance test for real-time PCS process under MRG-realtime (MRG-R) environment is done at pcsrt6 system. The system is optimized following the official tuning guide of MRG-R A special test program of measuring response time is used for testing performance at control cycles of 1/10/20/50 us For a single isolated process, the MRG-R is capable of safely running control cycles up to 10 us Future plan : Test compatibility of the communication hardware (RFM or digitizer) Test performace of the new PCS process for NTM mode detection acquisition 14
Recommend
More recommend
