Performance in Booting Linux w/ Device Tree via U-Boot out of JFFS2 on NAND

Performance in Booting Linux w/ Device Tree via U-Boot out of JFFS2 on NAND

[Grant Erickson]

I am continuing some experiments in booting Linux w/ a flattened device tree
via u-boot (1.3.2-rc3) from JFFS2 on NAND on an AMCC "Haleakala" board and
am curious if anyone has come up with some quantitative performance
characterizations of the various options (in all cases, u-boot lives on NOR
flash). The options I am evaluating are:

1) Put uImage and haleakala.dtb in their own "raw" NAND slices and boot with
   u-boot nand commands:

    static struct mtd_partition nand_parts[] = {
        {
            .name   = "kernel",
            .offset = 0,
            .size   = 0x0400000
        },
        {
            .name   = "fdt",
            .offset = 0x0400000,
            .size   = 0x0010000
        },
        {
            .name   = "root",
            .offset = 0x0410000,
            .size   = 0x3BF0000
        }
    };
    
    => nand read.i 200000 0 400000
    => nand read.i 400000 400000 10000
    => setenv bootargs ${bootargs} console=ttyS0,${baudrate}
    => setenv bootargs ${bootargs} root=/dev/mtdblock9 rootfstype=jffs2
    => bootm 200000 - 400000

Qualitative performance: Nearly instantaneous.

As expected, in this case the qualitative, subjective time to seeing "Linux
version 2.6.25-rc3-00951-g6514352-dirty ..." is nearly instantaneous.

2) Put uImage and haleakala.dtb as files in /boot in the ~12 MB JFFS2 root
   file system image in the ~60 MB "root" NAND slice and boot with u-boot
   fsload commands:

    => fsload 200000 boot/uImage
    => fsload 400000 boot/haleakala.dtb
    => setenv bootargs ${bootargs} console=ttyS0,${baudrate}
    => setenv bootargs ${bootargs} root=/dev/mtdblock9 rootfstype=jffs2
    => bootm 200000 - 400000

2a) With CFG_JFFS2_SORT_FRAGMENTS enabled.

Qualitative performance: Takes the better part of 30-35 minutes.

As expected with the in-documentation warnings about
CFG_JFFS2_SORT_FRAGMENTS and looking at the code in
u-boot/fs/jffs2/jffs2_nand_1pass.c, the qualitative, subjective time to
seeing the Linux version banner is slow, slow and slow.

2b) With CFG_JFFS2_SORT_FRAGMENTS disabled.

Qualitative performance: Takes about 30 seconds to two minutes.

3) This is a hybrid approach that I am setting up right now and is where I
am curious if anyone has done plots of fsload time on JFFS2 + NAND relative
to file system size.

Here, we use a separate 4 MB "/boot" JFFS2 file system for uImage and
haleakala.dtb files and a 60 MB "/" JFFS2 file system for the root file
system.

    static struct mtd_partition nand_parts[] = {
        {
            .name   = "boot",
            .offset = 0,
            .size   = 0x0400000
        },
        {
            .name   = "root",
            .offset = 0x0400000,
            .size   = 0x3C00000
        }
    };

    => fsload 200000 uImage
    => fsload 400000 haleakala.dtb
    => setenv bootargs ${bootargs} console=ttyS0,${baudrate}
    => setenv bootargs ${bootargs} root=/dev/mtdblock9 rootfstype=jffs2
    => bootm 200000 - 400000

3a) With CFG_JFFS2_SORT_FRAGMENTS enabled.

Shouldn't be necessary since the /boot file system would only ever be
accessed read-only and updated by nandwrite, not individual file updates.

3b) With CFG_JFFS2_SORT_FRAGMENTS disabled.

Qualitative performance: TBD <= 2b

Thanks,

Grant Erickson

Re: Performance in Booting Linux w/ Device Tree via U-Boot out of JFFS2 on NAND

[Grant Erickson]

On 3/6/08 9:30 AM, Grant Erickson wrote:
> I am continuing some experiments in booting Linux w/ a flattened device tree
> via u-boot (1.3.2-rc3) from JFFS2 on NAND on an AMCC "Haleakala" board and am
> curious if anyone has come up with some quantitative performance
> characterizations of the various options (in all cases, u-boot lives on NOR
> flash). The options I am evaluating are:
> 
> 1) Put uImage and haleakala.dtb in their own "raw" NAND slices and boot with
>    u-boot nand commands:
> 
> [ ... details omitted ... ]
> 
> Qualitative performance: Nearly instantaneous.
> 
> As expected, in this case the qualitative, subjective time to seeing "Linux
> version 2.6.25-rc3-00951-g6514352-dirty ..." is nearly instantaneous.
> 
> 2) Put uImage and haleakala.dtb as files in /boot in the ~12 MB JFFS2 root
>    file system image in the ~60 MB "root" NAND slice and boot with u-boot
>    fsload commands:
> 
> [ ... details omitted ... ]
> 
> 2a) With CFG_JFFS2_SORT_FRAGMENTS enabled.
> 
> Qualitative performance: Takes the better part of 30-35 minutes.
> 
> As expected with the in-documentation warnings about CFG_JFFS2_SORT_FRAGMENTS
> and looking at the code in u-boot/fs/jffs2/jffs2_nand_1pass.c, the
> qualitative, subjective time to seeing the Linux version banner is slow, slow
> and slow.
> 
> 2b) With CFG_JFFS2_SORT_FRAGMENTS disabled.
> 
> Qualitative performance: Takes about 30 seconds to two minutes.
> 
> 3) This is a hybrid approach that I am setting up right now and is where I am
> curious if anyone has done plots of fsload time on JFFS2 + NAND relative to
> file system size.
> 
> Here, we use a separate 4 MB "/boot" JFFS2 file system for uImage and
> haleakala.dtb files and a 60 MB "/" JFFS2 file system for the root file
> system.
> 
> [ ... details omitted ... ]
> 
> 3a) With CFG_JFFS2_SORT_FRAGMENTS enabled.
> 
> Shouldn't be necessary since the /boot file system would only ever be accessed
> read-only and updated by nandwrite, not individual file updates.
> 
> 3b) With CFG_JFFS2_SORT_FRAGMENTS disabled.
> 
> Qualitative performance: TBD <= 2b

For what it's worth, the results of (3b) above with a 4 MB "boot" JFFS2 file
system were the same as (2b) where "/boot" was just a subdirectory of the 12
MB (62 MB total NAND space) "/" JFFS2 file system:

In short, qualitative performance: Takes about 30 seconds to two minutes.

So, with CFG_JFFS2_SORT_FRAGMENTS disabled it would appear that fsload on
JFFS2 is O(1) with respect to one or all of: file system size, inodes or
dirents in the 4 MB to 64 MB range.

Regards,

Grant