Re: Grub2 EFI: Image loading from USB takes too long

[Aravind Srinivasan]

Tried the block IO patch. It is much better than the current default 
performance 
of 1.99. For the same initrd image (160M), loading from USB took approx 20 
secs. 
I also tried changing the value of GRUB_DISK_CACHE_BITS from 6 to 8, and load 
time improved to approx 12 secs.
Just to note, on Legacy Grub, this gets loaded almost in 3 secs.

Thanks,
Aravind


----- Original Message ----
From: Vladimir 'φ-coder/phcoder' Serbinenko <address@hidden>
To: The development of GNU GRUB <address@hidden>
Sent: Tue, April 5, 2011 12:36:35 AM
Subject: Re: Grub2 EFI: Image loading from USB takes too long

On 05.04.2011 08:32, Vladimir 'φ-coder/phcoder' Serbinenko wrote:
> On 05.04.2011 07:33, Aravind Srinivasan wrote:
>> include/grub/disk.h
>> <snip>
>> /* The size of a disk cache in sector units.  */
>> #define GRUB_DISK_CACHE_SIZE   8
>> #define GRUB_DISK_CACHE_BITS   3
>> <snip>
>> 
>> I tried changing this value to 8192 - and the time it took to load the same
>> image came down to 10 secs !
>> /* The size of a disk cache in sector units.  */
>> #define GRUB_DISK_CACHE_SIZE   8192
>> #define GRUB_DISK_CACHE_BITS   13
>> 
>> I am not sure of the reason for setting the cache size to a low value and any
>> possible issues in increasing this size. Appreciate any input on this.
> This value is a misnomer. It's not a cache size but a cache unit size. As 
>immediate effect it has that if one reads a single 512B, sector one reads an 
>entire block of 4K. Under normal conditions values of around 32K-128K 
>(6-8/64-256) would be optimal. It seems that some EFI implementations do a 
>"buffering" only to discard the buffered data. Some of the reports suggest 
>that 
>this braindamage is limited to disk io and doesn't happen with Block IO. 
>Values 
>you propose would result in 4M reads even when one needs just few sectors and 
>so 
>is suboptimal under normal conditions.
> I'll prepare few patches to investigate.
> After 1.99 the GRUB_DISK_CACHE_BITS will be increased to 6-8 range since, 
> it's 
>optimal and my AF support requires sector size to be smaller than cache unit 
>size.
> 
Attached a patch based on my AF patch which change to block IO.

-- Regards
Vladimir 'φ-coder/phcoder' Serbinenko