Performance for 64 bit hardware?

[Eddy G]

Hi all,

I'm new on this...

I plan to buy hardware to run upto about 10 Gig samples.
What would a typical hardware requirement be?
I want to run on Linux.

Performance:
Processor?
64 bits?
Or a quad core?

RAM?
Is the 4 Gig limit of the 32 bits machines a bottleneck for me or should that more than enough?

I'm still dubbing about running the Linux in VMware (so I will need some extra RAM for that) or make a dual boot machine.
I think VMWare maybe is less risky?
Any suggestions on that are also welcome!

Thanks,
Eddy

[sbenno]

go for 64bits
(though most CPUs are 64bit capable right now, but since under windows you need Vista 64 to use it in 64bit mode, most people are not aware that they do have a 64bit capable cpu).

although LS does not internally multi core optimized (except for the fact that the disk streaming thread and audio rendering thread are separate so you get some benefits from dual core
CPUs over single core ones), you can use jack-dmp and open multiple jack audio devices in LS which allows to use all your available CPU cores by spreading the instruments on multiple LS audio outs.

For large templates (like orchestral music) this is very useful as it allows good scalability in terms of polyphony.
Alex for example is using this method for his orchestral templates with great success (800 voices of polyphony using a quad cpu and multiple fast disks IIRC).

RAM ... as much as you can afford
on 32bit machines more than 4GB does not make sense and often it is hard to use more than 2-3GB due to the operating system needing a chunk of the address space.
So go for 64bit mainboard and cpu and chose one where you can upgrade to at least 8-16GB so even if you start with less RAM (ie 4GB) you can always add more later.
If you compile LS in 64bit mode it can take advantage of all the RAM you have installed, this means you can preload more instruments, increase the preload amount which helps to
increase polyphony and to minimize the stress you put on your HD.

VMWare (and similar virtual machines), don't touch them with a 10ft pole.
In order to provide low latencies LS needs to run on the bare metal, any VM or emulation layer inbetween will lower the overall performance (polyphony) and increase latency and cause audio dropouts.

[dahnielson]

Yes, go 64bit. And there's always an advantage of having multiple cores even if programs can't take advantage of them (by running a lot of parallel threads) at least Linux will balance instances of processes between the cores.

I would say go with as much RAM you can afford. I would probably get a less expensive processor (and upgrade later) and try max out the RAM from the start or (less waste) get a decent processor and carefully plan a head what RAM to get. E.g. with a mobo that has four DDR2 slots and can handle max 8 GB of RAM you would need 2GB modules and to take advantage of the dual-channel you need to install them in pairs (so you would need 2x2GB modules in a first step and then add 2x2GB when you upgrade -- and not buy inexpensive 1GB modules that you would need to replace when you decide to upgrade).

Also get a mobo with as little integrated stuff as possible, because you really want to have control over the IRQs and those buggers have a tendency to seize all the valuable IRQs (and IRQ sharing is a big no-no) and you end up disabling them anyway. Ideally you want a BIOS that let you reassign IRQs.

For serious use of LinuxSampler you need to run the Linux kernel patched with the RT-patches, set RLIMITS for RT-priority and run rtirq to sort out the IRQs (distributions like 64 Studio will do that for you):

* http://rt.wiki.kernel.org/index.php/RT_PREEMPT_HOWTO (general instructions)
* http://proaudio.tuxfamily.org/wiki/inde ... _RT_Kernel (instructions for Gentoo, but really good explanations)

A good audio interface (like M-AUDIO's "Delta" series -- keep a lookout for them in the classifieds for good bargains) is also highly recommended .

[Alex]

64bit
as much ram as you can put in.
Fast HD's.

Compile LS and libgig for optimised performance. (There are detailed instructions Christian put together in the main site.)

JACK comes in two versions, normal jack, and jackmp for multi processors.
I'm running a complete orchestra from one box, it works.

Enjoy yourself!

good luck,

Alex.

p.s. I'm planning out a monster box at the moment, with a tyan motherboard, and as much ram as i can get (64GB is the target.)
Remember that throughput for streaming is the target here, and is usually the bottleneck. Sata2 7200 HD's with 16 or 32mb cache are a good start point, but if you can stretch to 10000rpm HDs, then at least one of these for the system drive would keep things going nicely. (My intention, and humble opinion.)

As a complete aside to this, i'm also watching the development of solid state HDs, and will post here if i get the opportunity to try or use one of these. The technology just needs a little more time and maturity at the moment.

[dahnielson]

p.s. I'm planning out a monster box at the moment, with a tyan motherboard, and as much ram as i can get (64GB is the target.) Remember that throughput for streaming is the target here, and is usually the bottleneck. Sata2 7200 HD's with 16 or 32mb cache are a good start point, but if you can stretch to 10000rpm HDs, then at least one of these for the system drive would keep things going nicely. (My intention, and humble opinion.)

I guess you could also create a striped RAID for high throughput.

[sbenno]

I guess you could also create a striped RAID for high throughput.

While raid increases throughput it does not increase seek performance.
So for maximum polyphony the best is always to spread out your samples on multiple disks as evenly as possible which allows to maximize both data throughput and seek performance.

Of course solid state disks are the future, once the cost comes down as I wrote in an older post it will be more or less like having an extension of RAM, this means no RAM wasted for preload samples, no preload times, no limits to the number of instruments acessible at the same time. (though polyphony will be limited by the databus and cpu speeds).
As said we will return to "RAM sampling" with the difference that RAM will be almost unlimited and dirty cheap

[Eddy G]

Hi all,

Thanks a lot for sharing your experiences!

I want to try to build a sampled pipe organ.
If there are 40 ranks, and playing with 10 fingers and 2 feet, gives 40 * (10+2) := polyphony of 480, not?

I think I will go for 64 bits,
at least dual core,
at least space for 8 Gb RAM,
multiple fast SATA2 drives,
plain mobo (without on-board stuff)

Hope to get the time for all this...

Have fun,
Eddy

[Eddy G]

News about cheap SSD's. Sorry, it is in Dutch:
http://core.tweakers.net/nieuws/54237/o ... pslag.html

32 Gig for EUR 229
128 Gig for EUR 399

[Eddy G]

Finally we built us ourselves (my brother assembled it ) a nice machine: Q6700 machine on an Asrock board, 4Gb DDR2 mem (allowing 16Gb DDR3), and with the 64 Gb OCZ SSD with OpenSuse 11.0 installed on it. It boots within 20 secs. Not too bad I think.

Now I can focus on the soft side and see if I can get it running..!
Cheers, Eddy

[sbenno]

Hi Eddy, since you have the SSD could you run bonnie (disk I/O benchmark) on it ? it would be interesting to see how it would behave when using it using a sampler that does not any kind prebuffering
as I described in this thread above.

here you can find bonnie:
http://www.coker.com.au/bonnie++/

compile it and read the docs (it runs from the commandline, just use a filesize of 2GB or bigger if possible).
The most interesting part will be the seeks per second. (hopefully in the 0.x msec range)
it would be cool if you could post the output of bonnie here.

PS: what kind of polyphony do you get with LS from your box with SSD ?

cheers,
Benno