LynxOs 3.0.1 Performance - May the Best RTOS Win

LynxOS 3.0.1 PerformanceMay the Best RTOS Win . . .LynxOS Kernel Performance and SizingThis document provides run-time performance and memory image size benchmarks for the family ofLynxOS version 3.0.1 operating system kernels across four target microprocessor platforms.The most often requested measurements of• Context Switch• Interrupt Latencyare included herein.Furthermore, to provide example response times for typical systems service, the following measurementsare also included:• Driver Access Time• Memory Allocation/Deallocation• System Call Overhead• Semaphore• Getting Thread Priority• Mutex creation, locking and unlockingPerformance Test Suite UsedLynx has developed a performance test suite to simulate a real-world reactive system environment tomeasure meaningful performance for supported processor families.In order to simulate an event-driven real-time system with a realistic number of external interrupts drivingit, tests are conducted with five asynchronous interrupting devices:• network interface• console display• keyboard• disk controller• external timerThe LynxOS Scheduling Model - Threads and ProcessesA unique advantage of LynxOS lies in its POSIX-conformant scheduling and memory addressing models,specifically with reference to the basic LynxOS scheduling entity, the thread, and the basic virtualaddressing container, the process.Most competing RTOS products do not offer the flexibility and memory protection afforded by a completethread and process model, and rely on unprotected tasks running in a single flat address space. As such,these products' vendors quote performance data in terms of tasks.LynxOS Version 3.0.1 Performance Updated 01/28/99

The LynxOS performance data provided herein is couched in terms of threads, which for comparison's sakemay be compared to the tasks in other RTOS products.Context Switch TimeThis measurement represents the time it takes to switch between a currently-running thread and the nextscheduled thread. This measurement is an observed value collected on an uninstrumented and unmodifiedkernel through a combination of a user program and an installed dynamic driver, using timer accesses oneither side of the context switch; as such, it is empirically derived and not a theoretical nor syntheticmeasurement.Interrupt LatencyThis measurement is the minimum observed time taken by the LynxOS kernel to respond to an externalinterrupt and pass control to the installed interrupt service routine. This calculation serves system designersas a measure of the interrupt service overhead associated with responding to external events. In a normalLynxOS interrupt service design, actual event service processing does not occur in the ISR; instead, thekernel dispatches a LynxOS kernel thread in response to the interrupt, which can be prioritized andscheduled on a par with any other thread in the systemKernel threads allow interrupt routines to be very short and fast. An important key feature of LynxOS,kernel threads ensure predictable response even in the presence of heavy I/O.The Interrupt Latency calculation represented an observed result obtained by running a user thread incombination with an installed dynamic driver. The interrupt service routine of the driver samples the timerand this value is returned to the user process through shared memory.Driver Access TimeThe time it takes to access a device driver.Function Callsmalloc()free()getppid()sem_signal()getpriority()phtread_mutex_*()The time to allocate free memory. In this sampling, 512K is allocated,and the result is the average of multiple runs.The time to de-allocate memory back to the free memory pool. In thissample 512 KB are de-allocated, the result is the average of multipleruns.The time it takes to get the parent process ID. This measurement istaken with a cache forced to be dirty, so the result gives the overheadof a system call that has not previously been invoked and the codecached. The call getppid() was selected because it performs aminimum amount of processing (assignment statement and return).This Semaphore Signal is the time to post and respond to asemaphore.The time to get the priority of a thread.The time to create, lock, and unlock pthread_mutexes.LynxOS Version 3.0.1 Performance Updated 01/28/99

Summary of Interrupt and Task Response Timesfor Supported PlatformsThe following table lists performance data for the Motorola PowerPC 750, 604, Intel Pentium II, Motorola68060, and microSPARC II, each running LynxOS 3.0. All times are in microseconds.MotorolaPowerPC 750MotorolaPowerPC 604IntelPentium IIMotorola68060SMEµSPARC IIConfigurationComputer System cPCI 750 MTX PC-AT Clone MVME 177 SPARCStation 5Microprocessor PPC 750 PPC 604e Pentium II MC68060 µSPARC IICPU Clock Speed 233 MHz 200 MHz 333 MHz 50 MHz 85 MHzMain Memory 32 MB 32 MB 64 MB 32 MB 64 MBPerformanceContext Switch Time 2 2 2 21 41Interrupt Latency < 1 < 1 11 3 3Driver Access Time 1 2 3 10 9Memory Allocationmalloc()/free() 5 / 9 4 / 7 2 / 5 15 / 18 18 / 14System Call Overheadgetppid() 2 3 1 20 8Semaphore Signalsem_signal() 4 3 1 27 14Thread Prioritygetpriority() 5 5 3 36 11pthread mutex Callscreatelockunlock111111< 1< 1

LynxOS Kernel SizingThe LynxOS operating system offers the flexibility of configuration across three profile types• Ultra Embedded• Embedded POSIX• Full Featured RTOSwith numerous options for system services and extensions.As such, the LynxOS kernel family offers a broad range of price performance options, with targetfootprints ranging from 33K to 300K.The following table lists kernel sizes for a sampling of kernel profiles and configuration.CPUKernel ConfigurationCode SizeOnlyKernelImage SizeRun-timeFootprintUltra Embedded Profile 33K 37K 48Kx86 Embedded POSIX Profile 124K 147K 176KFull-featured LynxOS - Kernel-Only Profile 252K 263K 296KUltra Embedded Profile 64K 77K 97KPPC Embedded POSIX Profile 244K 289K 352KFull-featured LynxOS - Kernel-Only Profile 416K 451K 514KOther Performance DataA number of competing RTOS vendors continue to circulate a set of benchmarks performed in 1991 by ateam at the now defunct Super-Conducting Super-Collider Lab. Besides being over 7 years out of date, theperformance numbers in this document (known under various titles including "All Things Considered")were later retracted and republished by the benchmark's authors due to methodological issues. The actualresults of the SSCL RTOS evaluation proved most favorable to Lynx and indeed LynxOS was the chosenOS for the SCSL magnet control system.To learn more about this benchmark, read Real Considerations for Real-Time.LynxOS Version 3.0.1 Performance Updated 01/28/99