Re: [PATCH v4 0/7] lib/lzo: performance improvements

[Dave Rodgman <dave.rodgman@arm.com>]

Hi Markus,

What are your thoughts on how we should proceed with this patchset? You raised a few concerns in December - however, I'm 
not sure what further changes might be needed, if any. IMO this could be merged as it stands.

Regarding compatibility concerns: patchset v4 does not modify the behaviour of existing lzo. It introduces an 
independent algorithm (closely based on lzo); and also introduces some Arm performance benefits for existing lzo, 
without modifying the behaviour. So I don't see a compatibility risk.

You mentioned a crash on MIPS - do you have any details on this please? I have not seen any crashes in my testing so I'm 
not able to look into this without more data.

On 07/12/2018 3:54 pm, Dave Rodgman wrote:
> Hi Markus,
> 
> On 06/12/2018 3:47 pm, Markus F.X.J. Oberhumer wrote:> Request 3 - add lzo-rle; *NOT* acked by me
>  >
>  >     [PATCH 6/8] lib/lzo: implement run-length encoding
>  >     [PATCH 7/8] lib/lzo: separate lzo-rle from lzo
>  >     [PATCH 8/8] zram: default to lzo-rle instead of lzo
>  >
>  > It (1) silently changes the compressed data format
> 
> I'm not sure this is relevant: as a separate algorithm, there's no reason
> to retain the same format (although backwards compatibility can help with
> migration). If you know of a way to improve the compatibility aspect
> though, that would be great!
> 
>  > (2) crashes on MIPS,
> 
> Please could you provide more detail? I tested on x86-32, x86-64, arm,
> arm64 and big-endian MIPS64, but if there is an issue I missed I'd like to
> address it.
> 
>  > and (3) makes compression and decompression on typical data 10% slower on
>  > X86_64 with our internal benchmarks,
> 
> It is of course data-dependent. In my testing, as I mentioned previously, RLE
> without the other patches does regress slightly on high-entropy data, but
> offers a win on low-entropy data. For the right applications (e.g., zram),
> this makes it overall beneficial.
> 
>  > and (4) has to be carefully checked for buffer overflows.
> 
> This has been reviewed prior to sharing on LKML, and of course tested,
> but further review is of course welcome.
> 
>  > As a final comment, I question the quality your benchmarks - combining
>  > arch-related ARM64 improvements and algorithmic changes into one
>  > benchmark comparision is just unprofessional marketing.
> 
> I felt it was helpful to show overall performance with the complete patchset:
> this is what end-users experience. However, as you can see below, I also
> previously shared a summary of the two main components of the patchset to
> try and address this sort of concern:
> 
>  >> As a quick summary of the impact of these patches on bigger chunks of
>  >> data, I've compared the performance of four different variants of lzo
>  >> on two large (~40 MB) files. The numbers show round-trip throughput
>  >> in MB/s:
>  >>
>  >> Variant         | Low-entropy | High-entropy
>  >> Current lzo     |  242        | 157
>  >> Arm opts        |  290        | 159
>  >> RLE             |  876        | 151
>  >> Arm opts + RLE  | 1150        | 181

Regards

Dave