mpc 8260 performance..

mpc 8260 performance..

[gopi]

Hi..

  We are planning to use 8260 for our embedded system.  One of the
  job is to do routing of around 40 interfaces, with each interface
  at about 500Kbps duplex.

  We are planning to run it at about 260MHz and SDRAM, say about 256KB.

  Can 8260 take this load of routing without impacting the cpu usage
  too much(say max around 40% usage).  We can assume the MTU to be
  around 1500.

  Is anybody already using 8260 for a similar purpose?

  Are there any studies done on this/ any pointers on the web?

thanx in advance
gopi


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

RE: mpc 8260 performance..

[Allen Curtis]

>   We are planning to use 8260 for our embedded system.  One of the
>   job is to do routing of around 40 interfaces, with each interface
>   at about 500Kbps duplex.

What kind of interfaces? What kind of routing?

** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

RE: mpc 8260 performance..

[gopi]

interfaces: hdlc interfaces (external hdlc controller connected to 8260 which
            can write to SDRAM directly.  Hence 8260 wont be involved in packet copying).

routing:
                             +------------+
                  Ethernet   | 8260 based |hdlc
External world <===========> | embedded   |----- 40 systems
                             | system     |-----
                             |            |-----
                             +------------+

As you can see, the systems on hdlc are connected to external world
through the 8260 based embedded system.

At peak, 8260 will have to get/send of all the hdlc interfaces and
route them to external world over ethernet.

my main concern was that, at peak, what would be the 8260 resources usage.

hope this clarifies
thanx
gopi

PS: Please note that we are not trying to build a full fledged router.
It is only a means of reaching the right side systems from the
external world.

On Fri, 31 Jan 2003, Allen Curtis wrote:

> >   We are planning to use 8260 for our embedded system.  One of the
> >   job is to do routing of around 40 interfaces, with each interface
> >   at about 500Kbps duplex.
>
> What kind of interfaces? What kind of routing?
>


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

RE: mpc 8260 performance..

[Allen Curtis]

> interfaces: hdlc interfaces (external hdlc controller connected
> to 8260 which
>             can write to SDRAM directly.  Hence 8260 wont be
> involved in packet copying).
>
> routing:
>                              +------------+
>                   Ethernet   | 8260 based |hdlc
> External world <===========> | embedded   |----- 40 systems
>                              | system     |-----
>                              |            |-----
>                              +------------+
>
> As you can see, the systems on hdlc are connected to external world
> through the 8260 based embedded system.
>
> At peak, 8260 will have to get/send of all the hdlc interfaces and
> route them to external world over Ethernet.

The only processor overhead you should have is inspecting the HDLC packets
and repackaging them for routing purposes. Take some time and learn the
Ethernet and SKB portions of Linux. You should be able to preallocate all
your buffers. The HDLC ASIC will DMA into RAM, the 8260 will DMA out the
Ethernet packets. Achievable CPU utilization should be very low.


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

RE: mpc 8260 performance..

[Allen Curtis]

> Also, if you go to Motorola's web page you can download a CPM performance
> tool for the MPC8260.

True but he is using 1/10th of the CPM performance required to support just
the 2 FCC interfaces on the 8260....


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

RE: mpc 8260 performance..

[Steven Blakeslee]

Also, if you go to Motorola's web page you can download a CPM performance
tool for the MPC8260.

-----Original Message-----
From: Allen Curtis [mailto:acurtis@onz.com]
Sent: Friday, January 31, 2003 9:34 AM
To: gopi@india.tejasnetworks.com
Cc: linuxppc-embedded@lists.linuxppc.org
Subject: RE: mpc 8260 performance..



> interfaces: hdlc interfaces (external hdlc controller connected
> to 8260 which
>             can write to SDRAM directly.  Hence 8260 wont be
> involved in packet copying).
>
> routing:
>                              +------------+
>                   Ethernet   | 8260 based |hdlc
> External world <===========> | embedded   |----- 40 systems
>                              | system     |-----
>                              |            |-----
>                              +------------+
>
> As you can see, the systems on hdlc are connected to external world
> through the 8260 based embedded system.
>
> At peak, 8260 will have to get/send of all the hdlc interfaces and
> route them to external world over Ethernet.

The only processor overhead you should have is inspecting the HDLC packets
and repackaging them for routing purposes. Take some time and learn the
Ethernet and SKB portions of Linux. You should be able to preallocate all
your buffers. The HDLC ASIC will DMA into RAM, the 8260 will DMA out the
Ethernet packets. Achievable CPU utilization should be very low.


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

RE: mpc 8260 performance..

[gopi]

I understand a bit of Eth and skb code of linux.  One of my concerns
was interrupts.

In this case, the system can potentially be interrupted too
often(assuming an interrupt per frame received and transmitted,
it comes to nearly interrupt every few 10s of uSecs).  So, though the
packet processing doesnt take much time(involving movement of skbs
and modifying few fields), there is a hidden cost of context
swaps(kernel isr, bh, user space)

I am not sure of the effect of this on performance..

thanx
gopi

On Fri, 31 Jan 2003, Allen Curtis wrote:

> > interfaces: hdlc interfaces (external hdlc controller connected
> > to 8260 which
> >             can write to SDRAM directly.  Hence 8260 wont be
> > involved in packet copying).
> >
> > routing:
> >                              +------------+
> >                   Ethernet   | 8260 based |hdlc
> > External world <===========> | embedded   |----- 40 systems
> >                              | system     |-----
> >                              |            |-----
> >                              +------------+
> >
> > As you can see, the systems on hdlc are connected to external world
> > through the 8260 based embedded system.
> >
> > At peak, 8260 will have to get/send of all the hdlc interfaces and
> > route them to external world over Ethernet.
>
> The only processor overhead you should have is inspecting the HDLC packets
> and repackaging them for routing purposes. Take some time and learn the
> Ethernet and SKB portions of Linux. You should be able to preallocate all
> your buffers. The HDLC ASIC will DMA into RAM, the 8260 will DMA out the
> Ethernet packets. Achievable CPU utilization should be very low.
>


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

RE: mpc 8260 performance..

[Allen Curtis]

> In this case, the system can potentially be interrupted too
> often(assuming an interrupt per frame received and transmitted,
> it comes to nearly interrupt every few 10s of uSecs).  So, though the
> packet processing doesnt take much time(involving movement of skbs
> and modifying few fields), there is a hidden cost of context
> swaps(kernel isr, bh, user space)
>
> I am not sure of the effect of this on performance..

Sounds like a HW design issue you need to address. Perhaps it would be
better to interrupt on buffer full or some sort of data ready but timeout
condition. It is good to be concerned about IRQ but this does not affect CPM
performance...


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

Re: mpc 8260 performance..

[Dan Malek]

Allen Curtis wrote:

 > gopi wrote:
>>In this case, the system can potentially be interrupted too...


> Sounds like a HW design issue you need to address. Perhaps it would be
> better to interrupt on buffer full ....

High performance routers don't interrupt at all, typically have
carefully hand-tuned algorithms to maximize the use of processor cache
for instructions and ensure the data bus is aligned and bursting data.
Linux has been used in such designs, but the network drivers are
written specific to those applications.  The Linux network drivers are
examples of workstation connections and network layer interfaces, but
aren't suitable for these specialty applications.  They are suitable
for functional testing, but not for high performance routing.

An 8260 can process lots of communication data, obviously its performance
is totally dependent upon the software design and implemetation.


	-- Dan


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/