On Fri, 5 Jun 2015, Mark Hills wrote:
> On Fri, 5 Jun 2015, Alan Stern wrote:
>
> > On Fri, 5 Jun 2015, Mark Hills wrote:
> >
> > > I have a new USB 3.0 hard drive which appears to be throttled at
> > > approx 5Mbyte/sec.
> > >
> > > Is there a mystery kernel option that is required to enable full
> > > performance on these new drives?
> > >
> > > I have tried changing following:
> > >
> > > * Switching onboard USB 2.0 interfaces for a completely new USB 3.0
> > > interface (xhci_pci)
> > >
> > > * Enabling and disabling the "USB attached SCSI" CONFIG_USB_UAS
> > >
> > > But drive performance stays limited at 5Mbyte/sec.
> > >
> > > I've settled on using the USB 3.0 interface, and the dmesg (below) reports
> > > "SuperSpeed" and looks ok.
> [...]
> > By any chance, do you mount the filesystem with the "-o sync" option?
> > That will slow things _way_ down.
>
> It's not -- and also I can reproduce the results with no file system;
> just a 'dd' to /dev/sdc1.
>
> > If that's not the answer, try acquiring usbmon traces on the fast and
> > slow systems for comparison (see the instructions in
> > Documentation/usb/usbmon.txt.).
>
> Thanks for this -- I made some progress.
>
> I also tested reads this time which were ok (150Mbyte/sec), it's only
> writes that appear to be throttled.
>
> I have a sample of reads and writes from both systems below. I disabled
> 'uas' for the tests, droped caches etc. and used 'dd bs=1M' to the
> partition.
>
> The most striking thing is that the affected system consistently uses ~4KB
> URBs ("Len: 1"?) for writes.
>
> Whereas all the other requests seem to be a majority ~60KB URBs and "Len:
> 30"
What did you use for this monitoring? It isn't usbmon output.
> I can imagine this accounts for the lack of throughput, but not whether
> it's the cause or a symptom.
It must be a symptom -- that is, it can't be the ultimate cause. There
must be a reason why the write URBs are so small; the default limit is
120 KB.
> The lsusb output for the device is identical on both systems (see below)
>
> I imagine something lies in the code that generates the write requests,
> but I am starting from scratch here. I'd be greateful if you (or anyone
> else) has information that could point me to the next step.
I'm puzzled. Have you checked /sys/block/sdc/queue/max_sectors_kb?
Whatever the reason is, it doesn't seem to be connected with the USB
stack.
Alan Stern
--
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
