On 11 Apr 2014, at 16:06, Hans Petter Selasky <h...@selasky.org> wrote:
> On 04/11/14 06:33, Daniel O'Connor wrote:
>> Also, when it does connect at 5Gbps the speed seems quite slow - on my 
>> laptop (with USB controller VID 0x8086 PID 0x9c31 - Lynx point I think) I 
>> get 225MB/sec using libusb. On FreeBSD I get around 92MB/sec although only 
>> after lowering(!!) the amount read per transfer.
> 
> FreeBSD sets an IRQ latency of 125us, while the others use the default of 
> 62.5us. Are you double buffering the USB transfers? The IRQ latency can be 
> changed by editing a macro in the XHCI driver:
> 
> #define XHCI_IMOD_DEFAULT       0x000003E8U     /* 8000 IRQ/second */
> 
> At a rate of 225MB/s you need around 2x32Kbyte of buffer and you need to 
> avoid short transfers.

Interesting..

My test program looks like..
        for (i = 0; i < EP_FDNREQ; i++) {
            usb_xf[i].xf = libusb_alloc_transfer(0);
            usb_xf[i].idx = i;
            usb_xf[i].done = 0;
            usb_xf[i].submitted = 0;
            p = malloc(EP_FDXFAMT);
                
            libusb_fill_bulk_transfer(usb_xf[i].xf, h, EP_UDBUS, p, EP_FDXFAMT, 
usbcb, &usb_xf[i], 10000);
        }

I then submit all these and then have the call back log the speed (after N 
transfers) and reissue the request.
(I can send you the full code if you like)

I find that on OSX if I have..
#define EP_FDXFAMT      32768                   /* Number of bytes per tranfer 
*/
#define EP_FDNREQ       4                       /* Number of request to keep in 
flight */

I get 225MB/sec pretty much constantly, if I lower those values then the 
transfer rate is much lumpier.

With the same code I get 125MB/sec on FreeBSD.

I tried fiddling the numbers to get more but that seems to be the maximum.

Curiously if I increase the number of bytes per transfer to 64k the throughput 
drops to 86MB/sec.

Lowering it to 16k gives 125MB/sec, 8k gives 62MB/sec.

Finally, I ran systat -vmstat 1 while running the test and I see 4000 IRQ/sec 
on the xhci device, not 8000 as your comment above would suggest.

--
Daniel O'Connor software and network engineer
for Genesis Software - http://www.gsoft.com.au
"The nice thing about standards is that there
are so many of them to choose from."
  -- Andrew Tanenbaum
GPG Fingerprint - 5596 B766 97C0 0E94 4347 295E E593 DC20 7B3F CE8C






Attachment: signature.asc
Description: Message signed with OpenPGP using GPGMail

Reply via email to