Hi,
I have measured the throughput for different SRIO - direct IO transaction formats for TI DSP - 6618 using the bridge board setup.
Problem Statement:
Nread gives better throughput then Nwite and NwriteR.
Measurement Setup:
DSP on EVM 1 connected to DSP on EVM 2 using bridge board.
Measurement Procedure:
1. Send a doorbell packet from DSP on EVM_1 to DSP on EVM_2 and measure time T1 on DSP on EVM_2.
2. Send multiple packets N packets of transaction type say Nwite of one size say X bytes.
e.g. 50 transaction of say ( 1024 bytes - programming LSU with size of 1024 bytes - 50 times for averaging)
3. Send a doorbell packet from DSP on EVM_1 to DSP on EVM_2 and measure time T2 on DSP on EVM_2.
4. Measure performance based on following equations.
Byte count < 256
Performance with overhead = ((byte_count + 18) * 8 * sys frequency)/((T2-T1)/N)
Byte count >= 256
Performance with overhead = (((byte_count/256)*20 + byte_count) * 8 * sys clock freq)/((T2 - T1)/N)
5. Repeat the above procedure for different sizes of a transaction type and then for different transaction type.
Observation :
Nread gives better throughput then Nwite and NwriteR.
Can any one help me to find out what could be the reason for seeing such a kind of behavior. Or is there any issue with the throughput measurement technique being adopted for measurement.
LSU shadow registers are configured in the default mode.
Regards,
Chaitanaya