VME in CMS

Printable (html) Printable  (pdf)

1 Introduction 2 VME64x in CMS 3 VMEBridges     3.1 CAEN     3.2 SBS         3.2.1 Drivers         3.2.2 Old Drivers (SLC3)         3.2.3 Measurements

Performance Measurements The VME Hardware was a optical VME bridge from SBS (Model620) and a RAMix139 VME memory (32Mb). The PC was a PIII 800 MHz PC running Linux with a 2.4.x kernel. The measurements have been carried out with SBS driver and user library version v2p3p0 All measurements used the HAL library(version ver-03-05-test-04). Performance for single accesses This test was made running the "PerformanceTester" application contained in the HAL examples. It has been run one time with the dummy BusAdapter and one time with the BusAdapter for the VME Bridge. This allows to subtract the overhead due to the HAL library itself. Results using the VMEDummyBusAdapter =============================================================================== ATTENTION ==> the measured time values do not measure the process time but the total time elapsed during the measurement (laboratory-time) So, in order to get meaningfull measurements, be sure not to have other active processes on your machine. =============================================================================== (TimeoutException) Timeout during polling the item "rwFlagMask" 1000ms have passed wheras the timeout has been set to 1000ms 3680000 polls have been carried out (HardwareDevice::pollItem) read r/w-item with full mask : 2.57833us read r/w-item with full mask and offset : 2.95839us write r/w-item with full mask : 2.99179us write r/w-item with full mask and offset : 3.42508us read r-item with full mask : 2.53569us read r-item with full mask and offset : 2.91862us write w-item with full mask : 2.66218us write w-item with full mask and offset: 3.07601us unmaskedRead r/w-item with full mask : 2.66056us unmaskedRead r/w-item with full mask and offset : 2.90977us unmaskedWrite r/w-item with full mask : 2.65765us unmaskedWrite r/w-item with full mask and offset : 2.92356us setBit r/w-item with 1bit mask : 3.26963us setBit r/w-item with 1bit mask and offset : 3.71627us resetBit r/w-item with 1bit mask : 3.26833us resetBit r/w-item with 1bit mask and offset : 3.69591us setBit w-item with 1bit mask : 2.95803us setBit w-item with 1bit mask and offset : 3.35657us resetBit w-item with 1bit mask : 2.85544us resetBit w-item with 1bit mask and offset : 3.37863us =============================================================================== Results using the SBS620LinuxPCIBusAdapter Remarks: The items which need significant more time involve two VME accesses. (This is documented in the HAL documentation). In principle the SBS allows to map the VME Devices of the crate into memory. In this case single accesses can be performed much faster. However this technique has significant disadvantages for the software design. It is much harder to write a VMEBusAdapter for the HAL using memory mapping which works for an arbitrary crate configuration. Since there was no urgent need to improve the performance for single word VME accesses, such a busadapter has not been written. =============================================================================== ATTENTION ==> the measured time values do not measure the process time but the total time elapsed during the measurement (laboratory-time) So, in order to get meaningfull measurements, be sure not to have other active processes on your machine. =============================================================================== read r/w-item with full mask : 18.6139us read r/w-item with full mask and offset : 19.0851us write r/w-item with full mask : 33.4937us write r/w-item with full mask and offset : 34.2225us read r-item with full mask : 18.7048us read r-item with full mask and offset : 19.2273us write w-item with full mask : 18.5893us write w-item with full mask and offset: 18.9426us unmaskedRead r/w-item with full mask : 19.1327us unmaskedRead r/w-item with full mask and offset : 19.5662us unmaskedWrite r/w-item with full mask : 18.4104us unmaskedWrite r/w-item with full mask and offset : 19.0731us setBit r/w-item with 1bit mask : 34.2165us setBit r/w-item with 1bit mask and offset : 34.5748us resetBit r/w-item with 1bit mask : 33.7547us resetBit r/w-item with 1bit mask and offset : 34.6076us setBit w-item with 1bit mask : 18.9606us setBit w-item with 1bit mask and offset : 19.556us resetBit w-item with 1bit mask : 18.6855us resetBit w-item with 1bit mask and offset : 19.2885us =============================================================================== Performance for Block Transfers Measurement conditions: The measurements exclusively measure the block transfer performance of the VME bridge. Data blocks have been transferred for different sizes from the host via the VME bridge to the memory module and vice versa. The data has been generated with a random generator. The size of the data blocks has been varied from 256 bytes to 32Mb. Altogether 12 blocksizes where chosen. The data read back was always verified against the data previously written. No data error has been observed. The measurement at each data point was repeated 100 times. The mean values of the resulting transfer rates are summarized in the colored plot below. Additional plots have been produced, in order to show the distribution of the data throughput for the 100 measurements taken for at each block-size. Results Read and write performance as a function of the block size. Distribution of read throughput for fixed block sizes (100 Measurements). Distribution of write throughput for fixed block sizes (100 Measurements).

last changed: 2012-10-02 comments to: Christoph Schwick