Measurements of Some Basic Overheads

Some data related to overhead of empty loops, gettimeofday() and semaphore operations is provided below.

The table below shows some basic measurements. The columns of this table are briefly explained.

• Empty Loop: This is the measured cost of an empty C language loop using the standard "-O" optimization level. The loop has a structure such as:
  
  for (i = 0; i < N; i++){
  }
• Gettimeofday() overhead: Many measurement experiments as well as subelements make frequest use of gettimeofday(). The following (pseudo) code was used to measure this overhead.
  
  for (i = 0; i < N; i++){
     gettimeofday(result);
  }
• histogram record() overhead: Many measurement experiments as well as subelements make frequest use of record(). The following (pseudo) code was used to measure this overhead.
  
  for (i = 0; i < N; i++){
     record(result, hist);
  }
• Semaphore operation: The SPY memo lists a requirement for free semaphore acquisition time of <= 50 usec average and <=300 usec maximum. In our tests the semaphore overhead is measured as follows. Two processes, A and B, use two semaphores, sem_a and sem_b, to continuously signal to each other. Initially B starts and waits for sem_b to be signaled. A than starts and signals sem_b and than immediately waits for sem_a to be signaled. When B awakes it signals sem_a and again waits on sem_b. This cycle repeats "N" times. Thus one cycle consists of two semaphore operations each of which results in a context switch. The numbers shown are the total duration of the test divided by (2N). Thus the numbers in this column are the overhead of one semaphore operation and the resulting context switch to awaken the waiting process.
  
  The results under "RT" column were run with processes A and B running at the highest real-time priority supported by the OS. The results in the "non-RT" column were obtained by using the simple time sharing priorities. As noted Linux exhibited anomalous behaviour under RT priorities. We are currently investigating this further.
  
  It should be noted that the maximum values for this operation are under evaluation and not discussed herein. While mean value requirements are easily met the maximum value requirements will likely not be met except for specific real-time operating systems.

 

Platform	Empty Loop Overhead	Gettimeofday() overhead	record() overhead	Sema OP RT	Sema OP non-RT
Sun Ultra 5 (Solaris 2.7)	0.0037 usec	0.3521 usec		22.45 usec	22.71 usec
Dell Intel PII-350 MHz PC, Linux 2.2.5	0.0057 usec	1.2197 usec		7-13 usec	6.435 usec
Dell Intel PII-350 MHz PC, LynxOS 3.0.0	0.0057 usec	5.17 usec		12.74 usec	12.87 usec
Dell Intel PIII-450MHz PC, Timesys Linux/RT v1.1 - 2.2.14	.0045 usec	1.163 usec [RT-Clock] clock_gettime() 1.239 usec		Pending	Pending
PowerPC 604e 200 MHZ SBC, VxWorks 5.3.1	Pending	Pending		Pending	Pending
Micron Pentium 166, QNX RTP 6.00	0.0121 usec	4.29 usec		Pending	Pending
P-III 993MHz, Linux 2.4.10, gcc 3.0.3 256kb cache	0.00202 usec	0.4904 usec	0.2297 usec	Pending	Pending
AMD Athlon 1300MHz, Linux 2.4.10, gcc 3.0.3 256kb cache	0.00155 usec	0.3007 usec	0.1408 usec	Pending	Pending
Pentium IV 1600MHz, Linux 2.4.7-timesys-3.1.214 gcc 2.96 256kb cache (prio=500)	0.001258 usec	0.1432 usec	0.20975 usec	Pending	Pending
Intel(R) Xeon(TM) CPU 3.00Ghz 1024kb cache, Linux2.6.18 Uniproc gcc4.0.0 (Emulab PC3000, 1/24/2007)	0.000670 usec	0.575610 usec	0.096340 usec	Pending	Pending
Intel(R) Xeon(TM) CPU 3.00Ghz 1024kb cache, Linux2.6.12 Uniproc gcc4.0.0 (Emulab PC3000, 1/24/2007)	0.00069 usec	1.194970 usec	0.093630 usec	Pending	Pending
Intel(R) Xeon(TM) CPU 3.00Ghz 1024kb cache, Linux2.6.20-rt8-UNI w/ highres timer support. gcc4.0.0 (Emulab PC3000, 2/19/2007, gthaker)	0.00067 usec	0.350630 usec	0.095280 usec	-	-

JAVA Platforms
Platform	Empty Loop Overhead	System.time() overhead	Sema OP RT	Sema OP non-RT	Histogram record()
RPX-Lite w/JBED 1.1.4	1.084 usec	3.633 usec (com.jbed.runtime)	PENDING	PENDING	701.3 usec
Sun Ultra 5, Solaris 2.6	0.01484 usec	1.67 usec	PENDING	PENDING	2.5682 usec
Micron Pentium 166, QNX RTP 6.00	0.395 usec	86.6 usec	PENDING	PENDING	PENDING
P-III 993MHz, Linux 2.4.10 HotSpot Client VM (1.3.1-b24, mixed mode) [2001.10, GT]	0.0062 usec	0.655 usec (JNI call to gettimeofday)	PENDING	PENDING	0.28 usec

 

Return to benchmark menu