[karve]

Yet another viewpoint - you can guess I take an interest in this right ? :-) - I took a look at the pricing of a standard configuration for what's called an entry level system - a quad CPU Intel 6 core rack server system ( single mother board 4 CPU slots ) -HP (Hewlett-Packard) ProLiant Entry-level Server - 4 x Xeon X7460 2.66GHz - Rack ( RedHat Operating System)

It comes in at somewhere between 16,000 USD to 25,000 USD - ( and no drives ! and only 16gb memory and only DDR2 at that ) -

http://computers.pricegrabber.com/servers/HP-ProLiant-DL580-G5-Xeon-HC/m91898999.html

The wide spread in the price (16K to 25K ) alone makes me think hard about this. Then I checked the benchmarking at a standard place - the rating comes in at 18000. http://www.cpubenchmark.net/multi_cpu.html

Upthread I described my set up costs - for my overclocked, multiple core single slot motherboard setup - it comes in as base ( excluding overclocking and core restoration and excluding case, power supply and other cannibalized parts) benchmark at 5800 and cost 420 USD - ( add on the exclusions and call it 500 )

So to get to a 18,000 benchmark just multiple it by 3 - from the parallelism aspect, the same software I use is what would be there on the Proliant - I nor they have any special magic sauce. But to overestimate lets call it multiply by 4.

I'm kinda just doing the math and for a similar benchmark as the Proliant system - I'd come in at 2000 USD - 2000, not 20,000 ! ( and in my case you'd get redundancy, you'd get 48Gb extra memory and 6Terabytes of disk thrown in for free ! )

The numbers really are so far off - its weird ! Not that this surprises me - computing is like any other commercial field.

Once I understand this domain better, say 3 months time, I'd love to set up some head to head tests.

Let us know what you go for.

[JackieBadger]

We ended up going for 36GB Duel-quad core set-up which came to around $5K
It performs de novo transcriptome assemblies and SNP searches in under an hour, so suits our purposes very well.
For de novo whole genome assemblies one would need to purchase one of the more powerful cluster set-ups, I imagine.

The trickiest part of this whole equation, as mentioned, is in fact the bioinformatics. A true learning curve to say the least !

[frozenlyse]

It performs de novo transcriptome assemblies and SNP searches in under an hour, so suits our purposes very well.

Woah what organism are doing transcriptome assembly on in an hour? And how many reads?

[BaCh]

I'm kinda just doing the math and for a similar benchmark as the Proliant system - I'd come in at 2000 USD - 2000, not 20,000 ! ( and in my case you'd get redundancy, you'd get 48Gb extra memory and 6Terabytes of disk thrown in for free ! )

You are comparing apples and oranges there.

On the one hand: what you pay for on the bigger servers is the ability to have, e.g., anything between 1 and 8 CPUs with anthing between 2 and 8 cores each and up to ~400 GiB RAM (HP) or even 1 TiB RAM (Dell) *on* *one* *machine*. No cluster, no MPI, no whatever "cludge" to distribute your jobs across several machines, it's all there, ready to be easily accessed by a program.

On the other hand: a typical server with "home-grade" hardware has as first limitation the amount of RAM you can plug in. Motherboards for home use I know have a maximum of 6 memory slots which can be fitted with modules of 4 GiB each, adding up to 24 GiB. This is already plenty for quite some applications, and if one does not need more than a quad-core and this amount of RAM, then the machines will come in at something like 2 k€.

That being said: the configuration the original author now chose (dual-quad core with 36 GiB for USD 5k) seems quite reasonable to me. Not too shabby, and if it turns out it is too small, can certainly be upgraded.

B.

PS: Me too I'd be interested what organism you have to assemble transcript de-novo in an hour ... how many reads and which program?

[JackieBadger]

Woah what organism are doing transcriptome assembly on in an hour? And how many reads?

A couple of species of skate..... around 250MB for each transcriptome

[karve]

I think that's why I'm interested in some head to head tests - to see how the kludgey solutions of Map+Reduce or MPI work out against the oranges in this particular domain and get $/per solution unit metrics.

Those numbers and adoption rates will be part of the unfolding process - but I've seen some surprising, unfavored by some in the industry paradigm, technological solutions go mainstream - TCP/IP v OSI, networked PCs v server+dumb terminal based office systems, the ubiquitous 802 11.b/g ( now n ) based wi-fi versus whatever it was that Hughes Aircraft had, REST v SOAP and so on.

Adoption rates will tell the story of course.

[quantrix]

If you're only talking about analysing a relatively small number of experiments for a single lab then I don't think you need anything too fancy. The main requirement is for lots of memory (16GB is a pretty cost effective option these days). This will in turn require a 64 bit OS to make use of it.

Hi Simon,
Thanks for that informative reply. However, I had a query in that regard. When you say 16 GB of memory, do you mean 16 GB per core? (16X4=48GB per quad core) or 16 GB per processor? (i.e, 4GB per core)

Is there any advantage of putting something like 48 GB of RAM per CPU processor versus say 16 GB?
Thanks in advance...

[cpw]

I'm planning to set up a massive memory system for sequence assembly - due to performance/price ration I'd like to go with AMD's Opteron. Anyway, there are some benchmarks showing a better per-core performance of high-end Intel CPUs: This will of course make an impact if the application can't make use of multiple cores.
Is the pure CPU-Performance - means minor differences between in each case fast CPUs - such a significant parameter to regard for the performance of sequence assemblers or aren't most assemblers bound by memory?

Many thanks...

[JackieBadger]

I'm planning to set up a massive memory system for sequence assembly - due to performance/price ration I'd like to go with AMD's Opteron. Anyway, there are some benchmarks showing a better per-core performance of high-end Intel CPUs: This will of course make an impact if the application can't make use of multiple cores.
Is the pure CPU-Performance - means minor differences between in each case fast CPUs - such a significant parameter to regard for the performance of sequence assemblers or aren't most assemblers bound by memory?

Many thanks...

I imagine that people designing most next gen assemblers take this into consideration, and if anything have high upper memory bounds. With this in mind, the off-the-shelf packages I have used (CLC Genomics workbench, and SeqMan Pro) are certainly able to utilize a lot of memory, although I am not sure what their upper limit is. They are certainly designed for multi-core machines as you are able to maintain custom processor affinities throughout all iterations and thus maximize processing efficientcy (something that older programs e.g. Structure, are not able to do)

[karve]

I'm planning to set up a massive memory system for sequence assembly - due to performance/price ration I'd like to go with AMD's Opteron. Anyway, there are some benchmarks showing a better per-core performance of high-end Intel CPUs: This will of course make an impact if the application can't make use of multiple cores.
Is the pure CPU-Performance - means minor differences between in each case fast CPUs - such a significant parameter to regard for the performance of sequence assemblers or aren't most assemblers bound by memory?

Many thanks...

For my own interest purposes, can you be more specific about the Opteron ? After all, it comes in 12 core models ( 6176 SE ), 8 core and so on. I even have a now-ancient but only introduced 5 years ago dual-core Opteron 280 system still running :-)

Utilization of core, nodes, memory really is quite specific to the program ( and perhaps the user parameters ).

We have synthetic benchmarks for CPUs and Systems, there are suites of standard configurations of standard programs ( gaming titles, audio, video encoding, archiving etc ), there are open source benchmarks, there are industry-advocacy group audited standard ones.

I don't see too many domain specific benchmarks. Are there any for this field ?

Word of mouth and consensus views has its place in figuring it out but standardized runs ( same version, same parameters, same input data etc..) would really help in understanding this. Perhaps putting together a suite like this would help ? Then it could be downloaded by anybody, run and the results sent back and added to a database.

[cpw]

For my own interest purposes, can you be more specific about the Opteron ? After all, it comes in 12 core models ( 6176 SE ), 8 core and so on. I even have a now-ancient but only introduced 5 years ago dual-core Opteron 280 system still running :-)

Utilization of core, nodes, memory really is quite specific to the program ( and perhaps the user parameters ).

We have synthetic benchmarks for CPUs and Systems, there are suites of standard configurations of standard programs ( gaming titles, audio, video encoding, archiving etc ), there are open source benchmarks, there are industry-advocacy group audited standard ones.

I don't see too many domain specific benchmarks. Are there any for this field ?

Word of mouth and consensus views has its place in figuring it out but standardized runs ( same version, same parameters, same input data etc..) would really help in understanding this. Perhaps putting together a suite like this would help ? Then it could be downloaded by anybody, run and the results sent back and added to a database.

I will go with a 4-Socket System based on AMD's 6100 series (6174, 6176, 6180). Our main scope for this machine is de novo genome assembly, massive memory is therefore a very important factor, the planned system will have 512 GB. Both high-end Xeon (Nehaleem) and Opteron (Magny-Cours) use a non-uniform access pattern (NUMA) to make use of such an amount of memory. NUMA divides memory in closer/faster and (a bit) slower memory (provided by an other CPU) and the access strategies (topologies) differ a bit between AMD & Intel. But I think both systems are sufficiently 'coupled' that this will not make much of a difference. Hence the price/performance ration speaks clearly for the Opteron/Magny-Cours.