My next motherboard

Quote from nononsense:

... Till about 3 years ago, I only used scsi. Everything is now ATA100. All drives 200GB, 7200rpm. No SATA yet. It doesn't really bring you anything, except for the troubles of the past. I will use it when NCQ will be supported from all sides. (My current "#hdparm -tT" performance hovers between 55 and 60 Mb/s).
I have a keen eye on clustering but don't require it right now. I can less or more guess what nitro is trying to do from his posts long time ago. My approach is different. Clustering seems to be the way how state of the art engines are now making their marks. One day this will filter down, it is probably going already.
More...

We dont use SCSI for most of our workstations - the reasoning being that they are a more costly choice although the price difference has started to erode. We prefer to keep things on servers and , depending upon the server it may or may not have hardware raid support: the alternative being replication to a network disk array.

Clustering is really only needed if you absolutely need 100 percent availability or if you are working on problems that lend themselves to distributed solution across multiple nodes - typically loosely coupled. Out of the box clusters - be it Linux or $soft - really do not natively support other types of distributed problems and you will have to roll your own synchronization solution by altering the Linux kernel and making a few hardware changes if you want to work on those .. an excercise that is possible in linux because you have source and impossible on $soft.
 
Quote from linuxtrader:

You supply the knowledge: a programmer is not a trader. Highly skilled programmers and systems designers that are familiar with financial systems at the highest levels are available by the dozen. Even the most esoteric fianancial engineering applications are now commoditized.
More...
That’s what I thought you’d say. I describe how a function/algorithm should work. They write it. Yes, that might help somewhat, but it does not solve the real problem. Using a programmer to write my code may actually take more of my time to explain what to write and check the results than if I write it myself.

I was under the impression you were talking about programmers that could take a trading system model, understand it’s performance characteristics and apply it to other markets or vary the model in a productive way, on their own. Thus the burden is not just writing code, it is knowledge, analysis, coding and testing together. Specifically my work involves bringing in new market data, checking the data, defining indicator models, defining adaptive networks, trading models, writing efficient code, scheduling CPU jobs, analyzing results, trying to understand results, choosing systems for real-time deployment, monitoring real time results, and automating as much as possible. These are the things I need help on, and it is very difficult to find someone qualified, aaffordable and trustworthy to actually do all of that.

And I hope you dont actually think that programming and trading are one in the same job: they are not. If you think that they are then you are designing a recipe for disaster within your business.
More...
Yes, two totally different disciplines. Programming is only one of several skills that may help in create profitable trading systems.
 
Quote from nononsense:

Hi Prophet,
As I explained, since the P4 era I gave up on my earlier frustrating programming efforts for SMP machines. I currently run two P4's: 2.4 and 2.8GHz. I also keep an older 2*PIII mainly to test out linux communication things. As you see, nothing very planned, simply what the past left me. The speed of the P4's is...
More...
Thanks for your reply. What hardware would you purchase today? New or Used? One can get wholesale lots of older business PCs by the pallet pretty cheap:
http://cgi.ebay.com/ws/eBayISAPI.dll?ViewItem&category=40176&item=5127547318&rd=1

Regarding that mini-ITX cluster, it may be possible to get quantity or wholesale pricing of $50/node for 1GHz modules like this:
http://store.mini-box.com/ituner/ep10ne.html
Maybe the cost would be even lower if the manufacturer could be persuaded to not install most of the I/O logic on the MB. Most nodes would be driveless. I believe power consumption is less than 20W/node. The cluster would run off a 12V power bus, using DC-DC power supplies like these: http://store.mini-box.com/ituner/pw60caatxpos.html

Seems like there should be more solutions like http://www.orionmulti.com/ Anyone seen any low power AMD Althon mini-itx boards? How fast are Pentium-M CPUs at floating point? http://www.ibase-i.com.tw/mb890.htm
 
Quote from prophet:

Thanks for your reply. What hardware would you purchase today? New or Used? One can get wholesale lots of older business PCs by the pallet pretty cheap:
http://cgi.ebay.com/ws/eBayISAPI.dll?ViewItem&category=40176&item=5127547318&rd=1

Regarding that mini-ITX cluster, it may be possible to get quantity or wholesale pricing of $50/node for 1GHz modules like this:
http://store.mini-box.com/ituner/ep10ne.html
Maybe the cost would be even lower if the manufacturer could be persuaded to not install most of the I/O logic on the MB. Most nodes would be driveless. I believe power consumption is less than 20W/node. The cluster would run off a 12V power bus, using DC-DC power supplies like these: http://store.mini-box.com/ituner/pw60caatxpos.html

Seems like there should be more solutions like http://www.orionmulti.com/ Anyone seen any low power AMD Althon mini-itx boards? How fast are Pentium-M CPUs at floating point? http://www.ibase-i.com.tw/mb890.htm
More...
Hi Prophet,
If I had a choice, I wouldn't touch any of the older things. All I ever did was buy RDRAM memory at ebay. I think that the mini boxes are a better choice but I believe your price is without memory. This makes it expensive again comparing three of these to a say 3GHz P4 board.

Really, I am only starting to look at this. My distcc "cluster" is something you get with linux but all it does is help you compile.

I am awaiting the O'reilly book: "High Performance Linux Clusters with OSCAR, Rocks, OpenMosix, and MPI"
http://www.oreilly.com/catalog/highperlinuxc/
O'reilly has a pdf file with the chapter on "Management Software". The software problem is what I would like to find out now most about. (The image on the outside jacket looks like an impressive "cluster" :p )
In fact gentoo has already quite some ebuild section called "sys-cluster":
http://packages.gentoo.org/packages/?category=sys-cluster
I found "ganglia" of the O'reilly chapter in this section:
http://ganglia.sourceforge.net/
For an old-unix hand like you it may be a good preparation to fire up one of your rigs with linux, maybe in dualboot. (This is very easy these days, no more bootmagic or so. I boot everything, including XP with grub. Once you get the hang of it, you can't beat this.)

Further the IBM cluster projects seem very impressive but this is way beyond our present aims.

Be good,
nononsense
PS: This is C3:
http://www.csm.ornl.gov/torc/C3/index.html
It looks like I'll have to go back to my Python missionary act as all this seems to be written in Python. :cool:
 
Quote from prophet:

...Using a programmer to write my code may actually take more of my time to explain what to write and check the results than if I write it myself.
...
it is very difficult to find someone qualified, aaffordable and trustworthy to actually do all of that.
More...

Its does take more time - but coding and testing typically takes much more time than specifying a design.

At some point you need to decide where your time is best spent: Once you have a couple good system design techniques mastered and you know the details of efficient implementations, it is best to let others do the work and oversee the effort. Then you can concentrate on the trading models and trade management. If you try to do everything you typically get nothing done .....
 
Quote from linuxtrader:

Its does take more time - but coding and testing typically takes much more time than specifying a design.

[...]
More...
If the designer made a poor specification - this being not uncommon - the coding and testing times tend to infinity.

Foregive me linuxtrader, it's only a lill' nononsense-joke.

Be good,
nononsense
 
Quote from nononsense:

If the designer made a poor specification - this being not uncommon - the coding and testing times tend to infinity.

Foregive me linuxtrader, it's only a lill' nononsense-joke.

Be good,
nononsense
More...

.. or the flip side is that the coders finish really fast and show you that everything passed the tests .... except that the testing design was faulty and then the production systems performance time on key tasks tend towards infinity .....

I've seen both sides: the way to stop the one you mention is to ensure that the designer specifies the system limits and the key performance tests required to validate the design up front: these are verified before any coding begins .....
 
Quote from linuxtrader:

Its does take more time - but coding and testing typically takes much more time than specifying a design.

At some point you need to decide where your time is best spent: Once you have a couple good system design techniques mastered and you know the details of efficient implementations, it is best to let others do the work and oversee the effort. Then you can concentrate on the trading models and trade management. If you try to do everything you typically get nothing done .....
More...
I do everything myself. I get alot done.

True, I don't have a life...

nitro
 
You must log in or register to reply here.
Back
Top