Re: [libdc1394-devel] Frames dropped

[Damien D. <da...@do...>]

Hi Aravind,

On Fri, 2006-07-14 at 11:12 -0400, Aravind Sundaresan wrote:
> Hi,
> 
> We have a set up with 2 Indigita 4-port PCI-X Firewire cards on a Dell
> 690n. We have attached 7 Pixelink A-742 cameras to the cards. Each
> card has 4 independent channels for the 4 ports. The trigger is
> provided externally and is synchronised. We capture 900 frames per
> camera.
> 
> We experience dropped frames when we capture on all cameras with the
> following specifications. The ring buffer fill status is mostly zero.
> We do not write the images to disk.
> 
> Number of frames: 900
> Frame size : 512x512 (format 7)
> Frame rate : 60 fps
> Throughput/camera = 15 MBytes/sec
> Ringbuffer size : 20

If the card specs are true it should be OK.

> We are a bit puzzled where exactly the frame is dropped. It is usually
> in one of the 2 cases in our experience.
> - Shutter speed is such that the exposure time is >=
> 1/framerate-CONSTANT. This is not the case as the errors do not happen
> when there are fewer cameras.
> - Ring buffer is full. This doesn't seem to be the case either. In
> fact the max fill at any time is 4-5 frames.
> 
> We are unable to understand where the bottleneck causing frames to
> drop occurs. Could it be happening during the DMA transfer? In this
> case, do we need to tune anything at the DMA level.
> 
> We experience no frame drops if we capture with
> - 6 cameras (3 on each port).
> 
> We experience frame drops if we capture with
> - 6 cameras (4 on one port, 2 on another, images are not written to hard disk)
> - 6 cameras at 30 fps (3 on each port, images _are_ written to hard disk)

This may indicate that 4 cameras for one port is too much. It's a bit
strange, but who knows? Try to identify which camera(s) is/are loosing
frames. If it's always from one of the 4 cameras connected to a single
card then the bottleneck may be at the card. What happens if you have 4
cameras on one card and nothing on the other?

> The processor is a dual core Intel(R) Xeon(TM) CPU 3.20GHz with 2MB
> cache. The secondary hard disk is essentially 3 SATA disks in RAID 0.
> We stagger writes across different cameras to hard disk with fsync(),
> so that the disk I/O is steady. We also set scheduler to real time
> using sched_setscheduler().

I have no experience with RT so you're on your own here ;-)

> We would be grateful if anybody has any insight into this problem. If
> you need more details about the hardware set up, I'll be glad to mail
> them to you.
> 
> The Indigita QuikFire iDT804PCI host adapter has separate 1394 host
> chip for each port, enabling us to take advantage of "the PCI-X bus
> bandwidth of 4 concurrent 1394 bus streams through a single card". The
> indigita page is down, but the cached page is available at
> http://72.14.209.104/search?q=cache:wvo42l3g_1wJ:www.indigita.com/products/hostadapters/idt804pci.html+QuikFire+iDT804PCI&hl=en&gl=us&ct=clnk&cd=1

Still down, but the info of the cache shows that, in theory at least,
the card is not the issue. But the fact that 3+3 cameras is different
than 4+2 indicates that there is a card issue.

Damien

-- 
  _    Damien 'Takahara' Douxchamps, PhD
 ('-   Post-doctoral investigator
 //\   Image Processing Group, NAIST
 V_/_  http://damien.douxchamps.net/