Smalltalk › Pharo › Pharo Smalltalk Users

VW vs Pharo Performance

_‹ Previous Topic Next Topic _›

Classic

List

Threaded

5 messages Options

horrido

VW vs Pharo Performance

https://benchmarksgame-team.pages.debian.net/benchmarksgame/fastest/vw-pharo.html

Is VisualWorks really faster than Pharo? Why???

Sven Van Caekenberghe-2

Re: VW vs Pharo Performance

> On 31 Aug 2019, at 13:27, Richard Kenneth Eng <[hidden email]> wrote:
>
> https://benchmarksgame-team.pages.debian.net/benchmarksgame/fastest/vw-pharo.html
>
> Is VisualWorks really faster than Pharo? Why???

The Pharo code is single threaded, single core, while the VW code uses something called MatriX.VirtualMachines to create workers that run on multiple cores.

Presumably that makes a huge difference.

Also, such parallelism is easy for benchmarks that can be split in independent tasks, real world code is a completely different story.

Basically, this is not the same code that is being compared.

giorgiof

Re: VW vs Pharo Performance

hello,

Test seems old, both virtual machines (VW and Pharo) have upgraded in the meantime.

VW has always been faster that Pharo for what I know, up to recent time. Recently Pharo changed the vm technology so now it should be much faster, but I don'r know is faster that VW , so it would be nice to have a up to date comparison.

Related to Matrix, this is a nice way to have many smalltalk images running in parallel, but it doesn't seems the case of this test. It requires a specific way of writing the code, it is not as automatic as defining a piece of code "parallel". The thread model of VW and Pharo are still the same, green thread and (almost) single core systems

ciao

giorgio

On Sat, Aug 31, 2019 at 2:37 PM Sven Van Caekenberghe <[hidden email]> wrote:

> On 31 Aug 2019, at 13:27, Richard Kenneth Eng <[hidden email]> wrote:
>
> https://benchmarksgame-team.pages.debian.net/benchmarksgame/fastest/vw-pharo.html
>
> Is VisualWorks really faster than Pharo? Why???

The Pharo code is single threaded, single core, while the VW code uses something called MatriX.VirtualMachines to create workers that run on multiple cores.

Presumably that makes a huge difference.

Also, such parallelism is easy for benchmarks that can be split in independent tasks, real world code is a completely different story.

Basically, this is not the same code that is being compared.

Lorenzo

R: VW vs Pharo Performance

Ciao Giorgio,

vedo che sei sempre sulla braccia; perché non ci sentiamo per mettere insieme due vecchiaie?

Lorenzo

Da: Pharo-users [mailto:[hidden email]] Per conto di giorgio ferraris
Inviato: sabato 31 agosto 2019 17:48
A: Any question about pharo is welcome
Oggetto: Re: [Pharo-users] VW vs Pharo Performance

hello,

Test seems old, both virtual machines (VW and Pharo) have upgraded in the meantime.

ciao

giorgio

On Sat, Aug 31, 2019 at 2:37 PM Sven Van Caekenberghe <[hidden email]> wrote:

> On 31 Aug 2019, at 13:27, Richard Kenneth Eng <[hidden email]> wrote:
>
> https://benchmarksgame-team.pages.debian.net/benchmarksgame/fastest/vw-pharo.html
>
> Is VisualWorks really faster than Pharo? Why???

The Pharo code is single threaded, single core, while the VW code uses something called MatriX.VirtualMachines to create workers that run on multiple cores.

Presumably that makes a huge difference.

Also, such parallelism is easy for benchmarks that can be split in independent tasks, real world code is a completely different story.

Basically, this is not the same code that is being compared.

Ben Coman

Re: VW vs Pharo Performance

In reply to this post by Sven Van Caekenberghe-2

On Sat, 31 Aug 2019 at 20:37, Sven Van Caekenberghe <[hidden email]> wrote:

> On 31 Aug 2019, at 13:27, Richard Kenneth Eng <[hidden email]> wrote:
>
> https://benchmarksgame-team.pages.debian.net/benchmarksgame/fastest/vw-pharo.html
>
> Is VisualWorks really faster than Pharo? Why???

The Pharo code is single threaded, single core, while the VW code uses something called MatriX.VirtualMachines to create workers that run on multiple cores.

Presumably that makes a huge difference.

Also, such parallelism is easy for benchmarks that can be split in independent tasks, real world code is a completely different story.

Basically, this is not the same code that is being compared.

Thanks Sven for this insight. It wold not have occurred to me.

Looking at the results with new eyes, I notice its apparent in the sum activity of the four processors - averaging around 103% (including non-Pharo tasks) for Pharo testing

and much more for VisualWorks. Normalizing CPU usage I get these comparative results (smaller is faster VW).

0.86 fasta
0.93 binary-trees
1.02 fannkuch-redux
1.02 reverse-complement
1.08 pidigits
1.09 spectral-norm
1.44 n-body
1.49 k-nucleotide

So VW is 20% faster on some, 40% slower on others and most are pretty close.

Although such a Matrix feature would be advantageous for some domains - just need someone to invent similar for Pharo.

cheers -ben