[yocto] Performance improvements and machine build configuration

[Chris Tapp]

On 23 Oct 2012, at 19:45, Elvis Dowson wrote:

> Hi,
>      I noticed that between commits 
> 
> http://git.yoctoproject.org/cgit/cgit.cgi/poky/commit/?id=0260bb5c6978839c068007fcff2f704937805faf
> 
> and 
> 
> http://git.yoctoproject.org/cgit/cgit.cgi/poky/commit/?id=a3d5e9e6b7729319c518dcaf25bbe0643bfb25db
> 
> the build time has improved by around 7 minutes for my machine configuration, for building a core-image-minimal rootfs for the Xilinx ZC-702 FPGA with dual ARM Cortex A-9 CPUs.
> 
> commit id 0260bb5c6978839c068007fcff2f704937805faf        took 29 minutes
> commit id a3d5e9e6b7729319c518dcaf25bbe0643bfb25db  took 22 minutes
> 
> The machine configuration is an Intel i7 3770K over-clocked to 4.2GHz, with 16GB RAM at 1600Mhz, two 120GB SSDs configured into a striped disk array (Intel 330 series SSDs) with a write performance of 838MB/s and read performance of around 600MB/s, in RAID0 configuration, with a Corsair HT100 liquid CPU cooler keeping the CPU cool at around 52 degree centigrade during the build process. The motherboard is a gigabyte GA-Z77X-UP5TH
> 
> http://www.gigabyte.com/products/product-page.aspx?pid=4279#ov
> 
> This motherboard has a thunderbolt display port, so I can re-use my existing Apple Thunderbolt display. I've run Ubuntu 12.04.1 LTS and Ubuntu 12.10, and it appears to work after a few tweaks.
> 
> The only curious thing that I've noticed is that I don't see a large performance improvement using a standard 3TB Seagate Barracuda 7200 RPM HDD, and the two Intel Series 330 SSDs in a striped RAID0 configuration. The read (600MB/s) / write (838MB/s) figures are impressive, although I expected the read performance to be higher than write performance, as is normally with a single SSD. I'm using the motherboard's hardware RAID support on a 6GB/s SATA 3 port.
> 
> The 3TB HDD took the approximately 2 or 3 minutes longer than the 120GB x 2 RAID0 SSD configuration for commit id 0260bb5c6978839c068007fcff2f704937805faf (31 minutes vs. 29 minutes).
> 
> My local.conf parallelism settings were set to 6 threads for bitbake and make, for the quad-core (virtual 8 cpu cores)system.
> 
> Has anyone tried yocto builds with a 6-core, 8-core or 10-core Xeon processor system? How do those figures fare? I'm thinking my current bottleneck might be the CPU and not the HDD (?!), for the yocto build workloads, which I find curious and would like to confirm. 


I did quite a bit of experimenting with this a while back (similar spec, but with nearly 1000MB/s read/write SDD array). CPU was quad core with hyper-threading, so 8 virtual cores. I generally run with 16 threads, 16 parallel make as I find that the main performance hit is running out of stuff to keep all the cores busy.

Most of the time all 8 cores are maxed out, but around when the kernel gets built (and cross tools needed for it) I see the total CPU use drop to about 25%. This isn't because the system is I/O bound; it simply doesn't have enough tasks ready to run at that point in time.

I estimate that my 55 min build times would come down by 10 to 15 minutes if I could keep the CPUs busy (still, much better than the 10 hour build times on my previous system!).

I tried 'tinkering' with the run queue priority order, but all I proved was that inverting it (i.e. make the things that were previous given high-priority have low-priority) made no measurable difference to my build times! I'm trying not to think too much about that one ;-)

Chris Tapp

opensource at keylevel.com
www.keylevel.com