[Beowulf] Servers Too Hot? Intel Recommends a Luxurious Oil Bath

Andrew Holway andrew.holway at gmail.com
Mon Sep 10 05:09:31 PDT 2012


You make about as much sense as a Japanese VCR instruction manual!

2012/9/10 Vincent Diepeveen <diep at xs4all.nl>:
> You are rather naive to believe that public released numbers, other
> than total CPU's sold, will give you ANY data on
> what holds the biggest secrets of society. Secrecy is a higher
> priority here than for any
> military data.
>
> On military data it's tough to make money, as you risk getting a car
> accident.
>
> With trading secrets you can make BIG CASH and no one shoots you and
> no one ever got convicted for stealing something there AFAIK.
>
> What we DO know is a few hard facts that are not nice to hear.
>
> That is that each year inflation averaged over the past 70 years is
> historically 5%+, and with those populists that call
> themselves politicians, all of them overspending, that won't get
> lower coming few years. The average will rise
> of course.
>
> If you buy shares apple now. You get 1.7% dividend.
>
> That's LOWER than inflation.
>
> Historic dividend is quite some above 4%. Nearly 5%.
> Historic Indexation is another 7%+.
>
> So if you would have invested in DJI, say around 1935 and then in
> 2005 looked what you had made,
> you would've made an average of nearly 12% a year, dividend and
> indexation added up.
>
> If you do math with other indice, for example AEX of Netherlands, it
> started in 1983, and then calculate to 2005,
> you get actually nearly the same percentage of nearly 12%.
>
> Of course had you promised before 9/11 a percentage of 12% as a
> quant, you wouldn't even get hired, as they wanted 20%
> as an average a year, in order to pay for their own costs.
>
> Right now no one who is making those percentages based upon
> indexation and dividend is low.
>
> Several high frequency traders on other hand made an average profit
> of over 10% profit each year over the past 12 years.
> High frequency trading was just a few billion dollar worth in total
> earnings start this century.
>
> Right now even the guys who are 'long term' investing in the
> exchanges, they just trade futures and relative few
> options as a protection (if they're cheap).
>
> So entire market has moved nearly completely to derivatives right now.
>
> That's an amount of cash that if you write it out, you easily make a
> mistake in the number of zero's it has that number,
> that much it is.
>
> You really would be naive to believe they will give you or me ANY
> information.
>
> First of all no one has access to the bunkers of the traders where
> they are doing all the analysis.
> Not even the army, especially army guys, most intelligence dudes are
> known for stealing data and using for their own
> gain, as long as they don't get shot, especially military doesn't get
> in there.
>
> Speaking of your government. They don't even have the EQUIPMENT to
> verify what the traders are doing. the governments
> equipment works in the milliseconds still. Traders however measure in
> microseconds of course.
>
> Let's start with definitions.
>
> Something that calculates in a HPC center. You sure its name is
> 'server'?
>
> For companies that really will give you 0 information, it's rather
> difficult to count what they actually have.
>
> It's for example unclear how many of them are number crunching on
> gpu's. We can safely assume majority is not yet doing that.
> But i know of at least 1 hedgefund which has surprising amounts of
> guys hired who know something about gpgpu.
>
> Then 1+1 = 2 of course.
>
> Start looking at intel Xeon sales and IBM sales and realize how few
> HPC centers that do for governments nowadays lately got
> an 'upgrade'. Then come back with those sales numbers.
>
> Furthermore for such long term analysis, it doesn't matter if a
> machine sometimes goes broken. The google principle simply
> applies. Cheap number crunching hardware is what matters. It will
> need quite some bandwidth of course for the trading data
> and they need big storage.
>
> So looking at number of harddrives sold, and i'd look to SATA drives
> if i were you as those are cheaper, is a more interesting thing.
>
> The only fast low latency machines with fast cpu's you need is in the
> datacenter of the platforms and the exchanges.
>
> For the simulations it's all about having massive cpu power and there
> is a LOT of solutions to do that.
> The only thing we DO know is that the High Frequency Traders are
> still making good cash and nearly no one else is.
>
> So it's obvious everyone is moving there. If you do not build
> yourself a huge datacenter as a trader, then it's time to retire and
> enjoy your pension.
>
> The most important observation is that HFT can avoid worst case
> scenario's for you. Limit losses during big crashes.
>
> What is that worth to you?
>
> Will there be a new crash any soon knowing that markets don't like
> Europe overspend a tad under 3%, whereas
> USA is overspending 50%+ right now?  (spending divided by income).
>
> It seems that in 2008 democratic forces took care that the already
> volatile situation, which was there since 2007,
> that it crashed instantly.
>
> If republican forces want to do the same thing now, they have roughly
> 3 weeks left for a perfect timing to get their
> candidate elected.
>
> And for years to come after that, USA will be overspending too much.
> Will it get like Japan which has 230%+ debt and therefore an exchange
> that total imploded?
>
> No one knows.
>
> But we do know 1 hard fact and that's that only the high frequency
> traders made GOOD PROFITS in a SYSTEMATIC manner past 12 years
> and no one else did do that.
>
> By the way: the bigger the bank the slower they are to adapt to the
> new reality...
>
> On Sep 10, 2012, at 1:01 PM, holway at th.physik.uni-frankfurt.de wrote:
>
>>> A single tad larger trader typically has a 3000 machines and they're
>>> all equipped with a highspeed network.
>>
>> Europe installed over 2 million new servers in 2011 according to
>> Gartner.
>>
>> The Royal Bank of Scotland, one of the the most massive banks on the
>> planet, operates something less than 10k servers in their data synapse
>> grid but this is their modeling department and its not their high
>> frequency trading machines.
>>
>> Just because a bank has a lot of servers does not mean they all
>> need to be
>> located near the exchange. The Montecarlo stuff will help the HFT guys
>> devise triggers and rules which will be set up on their trading
>> systems.
>> The actual HFT system will probably only actually be a couple of
>> machines
>> in some basement somewhere in london.
>>
>>
>>>
>>> As slowly trader by trader is simply getting kicked bigtime by the
>>> traders that do have those 3000+ machines in their own datacenter,
>>> one by one they are all setting up their own calculation centers with
>>> more machines to do those simulations figuring out whether their
>>> package, based upon previous data, can make them big cash tomorrow on
>>> the exchanges.
>>>
>>> In most nations the high frequency traders will have the majority of
>>> the HPC hardware. More than government.
>>>
>>> On Sep 10, 2012, at 12:08 PM, holway at th.physik.uni-frankfurt.de
>>> wrote:
>>>
>>>> http://www.google.com/about/datacenters/inside/index.html
>>>>
>>>> High Frequency Trading is actually a tiny amount of the world wide
>>>> computer power. Far far less than HPC. This kit tends to be located
>>>> very close to the exchange anyway. and is generally uninteresting in
>>>> terms of green IT.
>>>>
>>>> I recently met with one of the technical directors of Deutsch Börse
>>>> which is the major central european stock exchange.. They are just
>>>> moving their central trading platform from some VMS clustery thing
>>>> (before my time) to a Intel / Infiniband / RedHat cluster of no more
>>>> than 200 nodes of which about half were gateway nodes (with
>>>> interface
>>>> cards for external connections to the finance houses). Its all small
>>>> time stuff.
>>>>
>>>> So please cease debasing these discussions with 'Iceland will never
>>>> have datacenters because you cant use it for HFT". Thou speaks from
>>>> thyne arse :) HFT is far smaller than HPC which is tiny anyway.
>>>>
>>>> Latency generally doesn't matter. I can ping www.whitehouse.gov in
>>>> 30ms. www.number10.gov.uk in 23ms and, kinda ironically, 43ms for
>>>> www.berlin.de (I live in Berlin).
>>>>
>>>> Also, really, this oil bath stuff will never happen in the main
>>>> stream. Some procurement process with buy it because of the numbers
>>>> and then regret it for all time in a semi public way. Then we can
>>>> all
>>>> forget it and get on with what we have always been doing; stealing
>>>> partitions from the office to direct the hot air out of the door
>>>> that
>>>> we have propped open with a chair.
>>>>
>>>> I think most of us have had to replace all of the memory in our
>>>> shiny
>>>> new cluster at some time or some other annoying task that involves
>>>> taking out every node and doing something annoying with a
>>>> screwdriver
>>>> and that sharp edge that kept slicing under your thumb nail. I
>>>> actually nearly chopped of the tip of my finger on a 9000RPM fan. It
>>>> was not a neat wound.
>>>>
>>>> Image doing that in rubber gloves with everything covered in goop.
>>>> Brilliant.
>>>>
>>>> ta for now.
>>>>
>>>> Andrew
>>>>
>>>>> -----BEGIN PGP SIGNED MESSAGE-----
>>>>> Hash: SHA1
>>>>>
>>>>> On 05/09/12 23:14, Robert G. Brown wrote:
>>>>>
>>>>>> I don't think they even bother with a proper enclosure per
>>>>>> motherboard.  Over the counter, commodity, cheap, almost hardware
>>>>>> agnostic.
>>>>>
>>>>> A Google talk I went to a few years back confirmed that's the case
>>>>> (or
>>>>> rather, lack of a case IYSWIW) for their nodes in their search
>>>>> clusters, but they did make the point they use more standard
>>>>> hardware
>>>>> for the services they make money from (like advertising) to get
>>>>> better
>>>>> reliability.
>>>>>
>>>>> Whether that's still the case, I don't know.
>>>>>
>>>>> cheers!
>>>>> Chris
>>>>> - --
>>>>>  Christopher Samuel        Senior Systems Administrator
>>>>>  VLSCI - Victorian Life Sciences Computation Initiative
>>>>>  Email: samuel at unimelb.edu.au Phone: +61 (0)3 903 55545
>>>>>  http://www.vlsci.org.au/      http://twitter.com/vlsci
>>>>>
>>>>> -----BEGIN PGP SIGNATURE-----
>>>>> Version: GnuPG v1.4.11 (GNU/Linux)
>>>>> Comment: Using GnuPG with Mozilla - http://www.enigmail.net/
>>>>>
>>>>> iEYEARECAAYFAlBNXbgACgkQO2KABBYQAh9qVgCeJnjX+yi0Lp5aq2W0dPD8Aovu
>>>>> a6kAnRiXr7y/AXbybjs5U549CpM/zx9Z
>>>>> =1K1G
>>>>> -----END PGP SIGNATURE-----
>>>>> _______________________________________________
>>>>> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin
>>>>> Computing
>>>>> To change your subscription (digest mode or unsubscribe) visit
>>>>> http://www.beowulf.org/mailman/listinfo/beowulf
>>>>>
>>>>
>>>> _______________________________________________
>>>> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin
>>>> Computing
>>>> To change your subscription (digest mode or unsubscribe) visit
>>>> http://www.beowulf.org/mailman/listinfo/beowulf
>>>
>>> _______________________________________________
>>> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin
>>> Computing
>>> To change your subscription (digest mode or unsubscribe) visit
>>> http://www.beowulf.org/mailman/listinfo/beowulf
>>>
>>
>>
>
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf


More information about the Beowulf mailing list