Hi Eric,
On Fri, 2004-11-05 at 07:36 -0800, Eric Blossom wrote:
> On Fri, Nov 05, 2004 at 10:44:36AM +0100, Axel Waggershauser wrote:
> > So, if the total throughput with 2 devices happens to be 32MB/sec as
> > well, I'd conclude that there is no controller available with a higher
> > throughput.
On Tuesday 30 November 2004 1:42 am, Feyd wrote:
> So lets have a list of properly aligned buffers, how can I submit it in one
> transfer? The usb_sg_init uses multiple urbs and thus multiple transfers,
> if I understand it correctly.
Driver protocols can call almost anything a "transfer",
but as
On Mon, 22 Nov 2004 12:52:05 +0100
Oliver Neukum <[EMAIL PROTECTED]> wrote:
> >
> > Whould an incorrect alignment/packetization cause a real error or only
> > hurt performance?
>
> A real error. Within a transfer you need to use maximum sized packets.
> If you don't the transfer will end prematu
Am Montag, 22. November 2004 12:45 schrieb Feyd:
> On Sat, 30 Oct 2004 11:07:40 -0400 (EDT)
> Alan Stern <[EMAIL PROTECTED]> wrote:
>
> > On Sat, 30 Oct 2004, Axel Waggershauser wrote:
> >
> > > I have no sufficient knowledge of the whole scatter-gather method but I
> > > thought version 2 (avoid
On Sat, 30 Oct 2004 11:07:40 -0400 (EDT)
Alan Stern <[EMAIL PROTECTED]> wrote:
> On Sat, 30 Oct 2004, Axel Waggershauser wrote:
>
> > I have no sufficient knowledge of the whole scatter-gather method but I
> > thought version 2 (avoiding copy_to/from_user with sg) would exactly be
> > the way to
On Friday 05 November 2004 07:36, Eric Blossom wrote:
> On Fri, Nov 05, 2004 at 10:44:36AM +0100, Axel Waggershauser wrote:
> > So, if the total throughput with 2 devices happens to be 32MB/sec as
> > well, I'd conclude that there is no controller available with a higher
> > throughput. I would hop
On Fri, Nov 05, 2004 at 10:44:36AM +0100, Axel Waggershauser wrote:
> Hi Eric,
>
> On Fri, 2004-10-29 at 12:51 -0700, Eric Blossom wrote:
> > On Fri, Oct 29, 2004 at 01:06:34PM +0200, Axel Waggershauser wrote:
> > > another question: Have you tried accessing two devices at the same time?
> > > If
Hi Eric,
On Fri, 2004-10-29 at 12:51 -0700, Eric Blossom wrote:
> On Fri, Oct 29, 2004 at 01:06:34PM +0200, Axel Waggershauser wrote:
> > another question: Have you tried accessing two devices at the same time?
> > If so, what was your total throughput per sec?
> I haven't, but I'll give it a try
a person at Cypress sent me tables that show that USB cards on PCI are
slower than on the motherboard, and as expected USB thru-put is slower
when more than 1 USB chip is in use at the same time.
Ted
On 11/2/2004 1:13 PM, Eric Blossom wrote:
On Tue, Nov 02, 2004 at 09:23:20PM +0100, Axel Waggers
On Tue, Nov 02, 2004 at 09:23:20PM +0100, Axel Waggershauser wrote:
> On Thu, 2004-10-28 at 13:25 -0700, Eric Blossom wrote:
> I have tried to reproduce your bandwidth results by patching the
> fx2_programmer
> (http://volodya-project.sourceforge.net/fx2_programmer.php) to use your
> fast_usb code
On Thu, 2004-10-28 at 13:25 -0700, Eric Blossom wrote:
> > I have to write a custom driver for a FX2 based camera module that
> > should be able to reach a bulk throughput of about 30 MBytes/sec from
> > the device to user space memory. The available CPU performance is less
> > than a 1GHz VIA EDEN
Am Sonntag, 31. Oktober 2004 00:13 schrieb Alan Stern:
> You missed the point -- the users of st and sg I was referring to are
> people using those drivers with USB-based devices. So the quality or
> quantity of SCSI adapter support is irrelevant to this discussion. (And
> remember that cdrecord
On Sat, 30 Oct 2004, Oliver Neukum wrote:
> Am Samstag, 30. Oktober 2004 19:50 schrieb David Brownell:
> > On Saturday 30 October 2004 09:50, Oliver Neukum wrote:
> > > 2. No iommu is rare
> >
> > I thought they were all rare ... most x86
> > systems don't have them, for starters.
> > (Though x86
Am Samstag, 30. Oktober 2004 19:50 schrieb David Brownell:
> On Saturday 30 October 2004 09:50, Oliver Neukum wrote:
> > 2. No iommu is rare
>
> I thought they were all rare ... most x86
> systems don't have them, for starters.
> (Though x86_64 has one.)
True.
Maybe the typical users of sg and st
On Saturday 30 October 2004 09:50, Oliver Neukum wrote:
> 2. No iommu is rare
I thought they were all rare ... most x86
systems don't have them, for starters.
(Though x86_64 has one.)
---
This SF.Net email is sponsored by:
Sybase ASE Linux Expr
Am Samstag, 30. Oktober 2004 17:07 schrieb Alan Stern:
> problem. Another part is that user-space pages may be located in high
> memory, not directly accessible to the kernel. Not all USB host
> controllers use DMA, and very few of them can go above the 4 GB cutoff.
> (However this issue doesn'
Am Samstag, 30. Oktober 2004 16:39 schrieb Axel Waggershauser:
> I have no sufficient knowledge of the whole scatter-gather method but I
> thought version 2 (avoiding copy_to/from_user with sg) would exactly be
> the way to go. As far as I understand it, any memory block in user-space
> may be at
On Sat, 30 Oct 2004, Axel Waggershauser wrote:
> I have no sufficient knowledge of the whole scatter-gather method but I
> thought version 2 (avoiding copy_to/from_user with sg) would exactly be
> the way to go. As far as I understand it, any memory block in user-space
> may be at most fragmented
On Fri, 2004-10-29 at 19:33 +0200, Duncan Sands wrote:
> On Friday 29 October 2004 18:50, Johannes Erdfelt wrote:
> > Are you talking about exporting scatter-gather via usbfs? Or having the
> > kernel avoid a copy from userspace by using scatter-gather internally?
> >
> > The former might be usefu
On Fri, Oct 29, 2004, Duncan Sands <[EMAIL PROTECTED]> wrote:
> On Friday 29 October 2004 18:50, Johannes Erdfelt wrote:
> > On Fri, Oct 29, 2004, Duncan Sands <[EMAIL PROTECTED]> wrote:
> > > Hi Johannes, maybe it should allocate multiple pages and then use scatter-gather
> > > io out of them. Th
On Fri, 29 Oct 2004 09:59:22 +0200, Duncan Sands <[EMAIL PROTECTED]> wrote:
> > > Most interesting. How many URBs do you use?
> >
> > It's configurable. I think I'm using 100 16KB URBs right now. That's
> > probably overkill, but our hardware can only tolerate about 200 us
> > worth of jitter b
On Fri, Oct 29, 2004 at 01:06:34PM +0200, Axel Waggershauser wrote:
> Hi Eric,
>
> another question: Have you tried accessing two devices at the same time?
> If so, what was your total throughput per sec?
>
I haven't, but I'll give it a try over the next couple of days and let
you know what happ
On Fri, Oct 29, 2004 at 09:59:22AM +0200, Duncan Sands wrote:
> > > Most interesting. How many URBs do you use?
> >
> > It's configurable. I think I'm using 100 16KB URBs right now. That's
> > probably overkill, but our hardware can only tolerate about 200 us
> > worth of jitter before it over/u
On Fri, Oct 29, 2004 at 10:07:33AM +0200, Axel Waggershauser wrote:
> Hi,
>
> 1. looks like every byte transfered with your "fast usb" lib is
> memcopied at least two times: in your fusb*.cc code and in the usbfs
> kernel code, right? Meaning worrying about sg-list "optimized" drivers
> is sort of
On Fri, Oct 29, 2004, Oliver Neukum <[EMAIL PROTECTED]> wrote:
> Am Freitag, 29. Oktober 2004 20:44 schrieb David Brownell:
> > On Friday 29 October 2004 11:13, Johannes Erdfelt wrote:
> > > > If usbfs instead grabbed a bunch of non-contiguous
> > > > pages, copied the data into them from
Am Freitag, 29. Oktober 2004 20:44 schrieb David Brownell:
> On Friday 29 October 2004 11:13, Johannes Erdfelt wrote:
> > >If usbfs instead grabbed a bunch of non-contiguous
> > > pages, copied the data into them from user-space, and then sent it (using
> > > scatter-gather io), then there is n
On Friday 29 October 2004 11:13, Johannes Erdfelt wrote:
> > If usbfs instead grabbed a bunch of non-contiguous
> > pages, copied the data into them from user-space, and then sent it (using
> > scatter-gather io), then there is no longer any memory pressure problem.
> > What's more, there woul
On Friday 29 October 2004 18:50, Johannes Erdfelt wrote:
> On Fri, Oct 29, 2004, Duncan Sands <[EMAIL PROTECTED]> wrote:
> > > It probably will be significantly less. We have buffer size issues with
> > > usbfs and libusb.
> > >
> > > Older versions of libusb used the synchronous bulk read/write c
On Fri, Oct 29, 2004, Duncan Sands <[EMAIL PROTECTED]> wrote:
> > It probably will be significantly less. We have buffer size issues with
> > usbfs and libusb.
> >
> > Older versions of libusb used the synchronous bulk read/write calls and
> > those were limited to a page size. We've since switche
Hi Eric,
another question: Have you tried accessing two devices at the same time?
If so, what was your total throughput per sec?
On Thu, 2004-10-28 at 13:25 -0700, Eric Blossom wrote:
> We built an FX2 based software radio peripheral and we routinely
> sustain 32MB/sec in either direction. CPU c
> 1. looks like every byte transfered with your "fast usb" lib is
> memcopied at least two times: in your fusb*.cc code and in the usbfs
> kernel code, right? Meaning worrying about sg-list "optimized" drivers
> is sort of unnecessary...
Memory bandwidth is usually much higher than USB bandwidth.
Hi,
On Thu, 2004-10-28 at 13:25 -0700, Eric Blossom wrote:
> We built an FX2 based software radio peripheral and we routinely
> sustain 32MB/sec in either direction. CPU consumption is minimal --
> on the order of 5% of a 1.4 GHz Pentium M. We did it all in user
> space using libusb, with an add
> > Most interesting. How many URBs do you use?
>
> It's configurable. I think I'm using 100 16KB URBs right now. That's
> probably overkill, but our hardware can only tolerate about 200 us
> worth of jitter before it over/under runs.
Do you ever have problems finding 100 * (4 contiguous pages)
> It probably will be significantly less. We have buffer size issues with
> usbfs and libusb.
>
> Older versions of libusb used the synchronous bulk read/write calls and
> those were limited to a page size. We've since switched to using
> asynchronous, URB based calls, which allow us buffer sizes
On Thu, 2004-10-28 at 16:18 -0700, Eric Blossom wrote:
> On Fri, Oct 29, 2004 at 12:27:45AM +0200, Oliver Neukum wrote:
> >
> > > We built an FX2 based software radio peripheral and we routinely
> > > sustain 32MB/sec in either direction. CPU consumption is minimal --
> > > on the order of 5% of
On Fri, Oct 29, 2004 at 12:27:45AM +0200, Oliver Neukum wrote:
>
> > We built an FX2 based software radio peripheral and we routinely
> > sustain 32MB/sec in either direction. CPU consumption is minimal --
> > on the order of 5% of a 1.4 GHz Pentium M. We did it all in user
> > space using libus
> We built an FX2 based software radio peripheral and we routinely
> sustain 32MB/sec in either direction. CPU consumption is minimal --
> on the order of 5% of a 1.4 GHz Pentium M. We did it all in user
> space using libusb, with an added layer on top that does "fast usb" by
> keeping multiple
> Hi,
>
> I have to write a custom driver for a FX2 based camera module that
> should be able to reach a bulk throughput of about 30 MBytes/sec from
> the device to user space memory. The available CPU performance is less
> than a 1GHz VIA EDEN. Is the libusb capable to provide this performance?
On Thu, Oct 28, 2004, Oliver Neukum <[EMAIL PROTECTED]> wrote:
> Am Donnerstag, 28. Oktober 2004 17:01 schrieb Alan Stern:
> > On Thu, 28 Oct 2004, Axel Waggershauser wrote:
> > > I have to write a custom driver for a FX2 based camera module that
> > > should be able to reach a bulk throughput of a
Am Donnerstag, 28. Oktober 2004 17:01 schrieb Alan Stern:
> On Thu, 28 Oct 2004, Axel Waggershauser wrote:
>
> > Hi,
> >
> > I have to write a custom driver for a FX2 based camera module that
> > should be able to reach a bulk throughput of about 30 MBytes/sec from
> > the device to user space me
On Thu, 28 Oct 2004, Axel Waggershauser wrote:
> Hi,
>
> I have to write a custom driver for a FX2 based camera module that
> should be able to reach a bulk throughput of about 30 MBytes/sec from
> the device to user space memory. The available CPU performance is less
> than a 1GHz VIA EDEN. Is t
Hi,
I have to write a custom driver for a FX2 based camera module that
should be able to reach a bulk throughput of about 30 MBytes/sec from
the device to user space memory. The available CPU performance is less
than a 1GHz VIA EDEN. Is the libusb capable to provide this performance?
If not, I gu
42 matches
Mail list logo
