File I/O Throughput Problem
 

I have an IBM x3650 M4 server (64-bit) running Ubuntu 13.04 Desktop (32-bit). I realize that this is a non-standard setup by it's being driven by other requirements. I am seeing slow file writes accross the board (to HDD, to USB, to network mount). The file writes seem to peek at 1.3Mb/s. The expected speed of my RAID setup is 180Mb/s.

Here's what I tried:

1. Boot Live Ubuntu 11.10 64-bit. The file writes are as expected, 180Mb/s.
2. Boot Live Ubuntu 13.04 32-bit. The file writes are slow, 1.3Mb/s.
2. Boot Live Fedora 19 32-bit. The files writes are again as expected, 180Mb/s.

At this point, it looks like something in the Ubuntu 13.04 32-bit kernel is not being configured correctly to work with the 64-bit hardware. Fedora 32-bit seems to work just fine.

Would anyone happen to know what could be causing this?

Sounds like a bug in the kernel, plain and simple. I'm willing to bet it has nothing to do with the CPU being capable of 64-bit addressing, as all modern CPUs are capable of that, and there's no such thing as "64-bit hardware" besides the CPU.

I believe Ubuntu 13.04 uses kernel 3.8.8 by default, while Ubuntu 11.10 uses 3.0.0. Fedora 19 seems to run either 3.9 or 3.10(?). If my theory is correct, the 64-bit Ubuntu 13.04 Live DVD should have the same performance issues.

It's difficult to say what what could be causing this without knowing a bit more about your setup and the tests you ran. Since you're seeing poor performance across the board rather than with a particular device, it's probably not an issue with a controller driver. At it happens both with local file systems and network mounts, bugs in file system drivers are also pretty much ruled out.

If you can't avoid using an affected kernel, check the logs for any errors. Run iperf tests to see if network performance in general is affected. Also, run dmesg and see what the kernel says about ACPI, PCIe and interrupts during boot.

If you have more than 4 GB RAM you might run into bounce buffer issues, where the OS is programming DMA with 32-bit physical addresses, then copying the data to/from high mem. Check CONFIG_HIGHMEM64G and CONFIG_ARCH_DMA_ADDR_T_64BIT.

That can certainly be an issue with 32-bit systems with more than 4 Gb memory, but I wouldn't expect it to cause a 99% decrease in performance.