> From: Alan Stern [mailto:stern@xxxxxxxxxxxxxxxxxxx] > Sent: Thursday, November 01, 2012 7:48 AM > > On Wed, 31 Oct 2012, Paul Zimmerman wrote: > > > > From: Michal Nazarewicz [mailto:mpn@xxxxxxxxxx] On Behalf Of Michal Nazarewicz > > > Sent: Wednesday, October 31, 2012 1:10 PM > > > > > > On Wed, Oct 31 2012, Paul Zimmerman wrote: > > > > Regarding that, would you guys accept a patch that adds the "buflen" > > > > module parameter to the mass-storage gadget? Without increasing the > > > > buffer length above 16K, it's impossible to approach the max throughput > > > > allowed by SuperSpeed hosts. > > > > > > I have no problem with that. But did you try setting up 4 buffers? you > > > can set number of buffers g_{file,mass}_storage uses with config > > > USB_GADGET_STORAGE_NUM_BUFFERS (or if you select USB_GADGET_DEBUG_FILES > > > you get a fsg_num_buffers module parameter). > > > > Yeah, I tried that, but it didn't have much effect. Whereas increasing num_buffers > > to 4 and buffer length to 128K gives a large improvement, from less than 200 MB/s > > to about 290 MB/s. > > No doubt this is a matter of interrupt latency. I don't know the > maximum throughput for SuperSpeed bulk data, so I'll guess it's about 8 > times that of high speed (500 MB/s vs. 480 Mb/s). High speed bulk data > can be transferred at rates up to 6.5 KB/uframe (although the values > one sees in practice are almost never that high). This means > SuperSpeed bulk data can be transferred at rates up to 52 KB/uframe. > Assuming we get completion interrupts only at microframe boundaries, > bulk data buffers should be at least 64 KB, and preferably larger. In our experience, 128 KB seems to be the sweet spot. With buffers larger than that you can get slightly better performance, but it's probably not worth the extra RAM unless you're trying to squeeze out the last drop of performance. -- Paul -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
