Forum Discussion

Altera_Forum's avatar
Altera_Forum
Icon for Honored Contributor rankHonored Contributor
16 years ago

Poor performance with PCIe and DMA on Linux

Hello all,

I built a evaluation system on the Arria II dev board that transports two similar datastreams to a linux host so that each stream can be read from a different device file.

The sopc system uses a dma controller that reads data from a fifo. The fifos are located in a custom component. The fifo is twice the maximum dma transfer length. The fifo generates a waitrequest if it's less than half full.

Each stream has a datarate of approx. 62.5 Mb/s, so I need 125 Mb/s what should be possible with a PCIe-x1 port.

If I measure the throughput with `dd`, I reach 40.1 Mb/s reading one stream at a time.

I realized that waitrequest from the PCIe-Tx-port throttles the data flow significantly.

My system is homogenous with 32 bit datawidth.

Does somebody has a idea how I can speed up my data transfers?

Thanks in advance.

Regards,

Frederik

1 Reply

  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    16 years ago

    Hello again,

    I did more tests.

    I doubled the datawidth from 32 bit to 64 bit. The fifos write doublewords so the throughput is doubled in theory, because the pcie-tx-port ist 64 bit wide.

    In fact the throughput isn't better now.

    If I increase the DMA-block-length, the whole card freezes. The PCIe-interface sets waitrequest and stays in an that state. The card does not react to any command until I reset it. If I put debugging messages into my driver's code, the block length can be bigger, because it takes time until the next dma-request.

    This issue occured with singleword transfers, too.

    Does anybody has an idea?

    Regards,

    Frederik