SRIO DirectIO testing rate result - Processors forum - Processors

0 Ganapathi Dhandapani95 over 12 years ago

Hi Asma,

SWRITE and NWRITE approach theoretical limits with large packets. However, since reads require waiting for a response, their throughput is lower and is highly dependent on implementation. NWRITE_R throughput is also lower since it requires a response. Throughput improves for all transfer modes as the packet size increases.

Please refer "SRIO Throughput" section on below links.

http://www.ti.com/lit/spraay0

http://www.ti.com/lit/sprabk5

Thanks,

0 asma elmaizi over 12 years ago in reply to Ganapathi Dhandapani95

Prodigy 150 points

Hy Ganapathi,

Thanks for your reply , I have verified both documents and it confirms the results that I found .

I understand the fact that Nread is slower than Nwrite because it wait for a response but my issue is :

why when we use C6474 until (4KB) which is the maximum size of a transaction Nread in C6474 is slower compared to Nwrite but in the case of C6678 when I did the tests from the value (8KB) to 1MB which is the maximum size Nread and Nwrite got an equal throughput .

Thanks.

Regards,

Asma

C6474

C6678:

0 Ganapathi Dhandapani95 over 12 years ago in reply to asma elmaizi

TI__Mastermind 28085 points

Hi Asma,

C6474 SRIO (rev 1.3 complaint) and C6678 SRIO (rev 2.1 complaint).

If you want to improves throughput for all transfer modes as the increases of packet size. C6474 SRIO Byte_count support up to 4K bytes only. But C6678 SRIO Byte_count support up to 1MB bytes. C6678 higher packet size transfer only got an equal throughput on Nread and Nwrite, small packet size transfer are not getting equal throughput.

Thanks,

0 asma elmaizi over 12 years ago in reply to Ganapathi Dhandapani95

Prodigy 150 points

Hi Ganapathi,

Thank you for your reply, this answered my question, it was very helpful for me.

Thanks.

Regards,

Asma

0 tscheck over 12 years ago in reply to Ganapathi Dhandapani95

TI__Mastermind 23525 points

In addition to what Ganapathi mentioned, I just want to point out that the LSU designs between the C64x and C66x devices are completely different. It is not a function of the SRIO spec compliance, it is simply implementation improvements that allow better throughput on C66x for a given datarate. Each LSU in C64x does one packet at a time, so for NREAD/NWRITE_R transactions that are greater than 1 packet, it will send one packet then wait for the response, then send another packet, wait for the response, etc. Each LSU on C66x, can track up to 16 outstanding transactions, so for transfers greater than 1 packet, it will send up to 16 packets and as soon as it receives the 1st response, it sends the 17th packet. In this way we make a lot better utilization of the link bandwidth.

Regards,

Travis

0 asma elmaizi over 12 years ago in reply to tscheck

Prodigy 150 points

Hi Travis ,

Thank you so much i appreciate your clarification , this answer my question clearly .

Regards,

Asma