hadoop-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Michael Segel <michael_se...@hotmail.com>
Subject Re: Hardware Selection for Hadoop
Date Tue, 07 May 2013 12:53:02 GMT
I wouldn't.

You end up with a 'Frankencluster' which could become problematic down the road. 

Ever try to debug a port failure on a switch? (It does happen and its a bitch.) 
Note that you say 'reliable'... older hardware may or may not be reliable.... or under warranty.
(How many here build their own servers from the components up?  ;-) 

I'm not suggesting that you go out and buy a 10 core cpu, however, depending on who you are,
and what your budget is... it may make sense. o 
Even for a proof of concept. ;-) 

While we have a rough metric on spindles to cores, you end up putting a stress on the disk
controllers. YMMV.

As to spending $$$ on hardware for  a PoC, its not only relative... but also what makes you
think this is the first PoC and only PoC he's going to do? The point is that hardware is reusable
and it also sets a pattern for what the future cluster will look like. After this PoC, why
not look at Storm, Mesos, Spark, Shark, etc... 

Trust me, as someone who has had to fight for allocation of hardware dollars for R&D...
get the best bang you can for your buck.

HTH

-Mike

On May 6, 2013, at 5:57 PM, Patai Sangbutsarakum <Patai.Sangbutsarakum@turn.com> wrote:

> I really doubt if he would spend $ to by 10 cores on a die CPU for "proof of concept"
machines.
> Actually, I even think of telling you to gathering old machines (but reliable) as much
as you can collect.
> Put as much as disks, Ram you can. teaming up NIC if you can, and at that point you can
proof your concept up to certain point.
> 
> You will get the idea how is your application will behave, how big of the data set you
will play with
> is the application cpu or io bound, and from that you can go out shopping buy the best
fit server configuration. 
> 
> 
> 
> On May 6, 2013, at 4:17 AM, Michel Segel <michael_segel@hotmail.com> wrote:
> 
>> 8 physical cores is so 2009 - 2010 :-)
>> 
>> Intel now offers a chip w 10 physical cores on a die. 
>> You are better off thinking of 4-8 GB per physical core. 
>> It depends on what you want to do, and what you think you may want to do...
>> 
>> It also depends on the price points of the hardware. Memory, drives, CPUs (price
by clock speeds...) you just need to find the right optimum between price and performance...
>> 
>> 
>> Sent from a remote device. Please excuse any typos...
>> 
>> Mike Segel
>> 
>> On May 5, 2013, at 1:47 PM, Ted Dunning <tdunning@maprtech.com> wrote:
>> 
>>> 
>>> Data nodes normally are also task nodes.  With 8 physical cores it isn't that
unreasonable to have 64GB whereas 24GB really is going to pinch.
>>> 
>>> Achieving highest performance requires that you match the capabilities of your
nodes including CPU, memory, disk and networking.  The standard wisdom is 4-6GB of RAM per
core, at least a spindle per core and 1/2 to 2/3 of disk bandwidth available as network bandwidth.
>>> 
>>> If you look at the different configurations mentioned in this thread, you will
see different limitations.
>>> 
>>> For instance:
>>> 
>>> 2 x Quad cores Intel
>>> 2-3 TB x 6 SATA         <==== 6 disk < desired 8 or more
>>> 64GB mem                <==== slightly larger than necessary
>>> 2 1GBe NICs teaming     <==== 2 x 100 MB << 400MB = 2/3 x 6 x 100MB
>>> 
>>> This configuration is mostly limited by networking bandwidth
>>> 
>>> 2 x Quad cores Intel
>>> 2-3 TB x 6 SATA         <==== 6 disk < desired 8 or more
>>> 24GB mem                <==== 24GB << 8 x 6GB
>>> 2 10GBe NICs teaming    <==== 2 x 1000 MB > 400MB = 2/3 x 6 x 100MB
>>>  
>>> This configuration is weak on disk relative to CPU and very weak on disk relative
to network speed.  The worst problem, however, is likely to be small memory.  This will likely
require us to decrease the number of slots by half or more making it impossible to even use
the 6 disks that we have and making the network even more outrageously over-provisioned.
>>>  
>>> 
>>> 
>>> 
>>> On Sun, May 5, 2013 at 9:41 AM, Rahul Bhattacharjee <rahul.rec.dgp@gmail.com>
wrote:
>>> IMHO ,64 G looks bit high for DN. 24 should be good enough for DN.
>>> 
>>> 
>>> On Tue, Apr 30, 2013 at 12:19 AM, Patai Sangbutsarakum <Patai.Sangbutsarakum@turn.com>
wrote:
>>> 2 x Quad cores Intel
>>> 2-3 TB x 6 SATA
>>> 64GB mem
>>> 2 NICs teaming
>>> 
>>> my 2 cents
>>> 
>>> 
>>> On Apr 29, 2013, at 9:24 AM, Raj Hadoop <hadoopraj@yahoo.com>
>>>  wrote:
>>> 
>>>> Hi,
>>>>  
>>>> I have to propose some hardware requirements in my company for a Proof of
Concept with Hadoop. I was reading Hadoop Operations and also saw Cloudera Website. But just
wanted to know from the group - what is the requirements if I have to plan for a 5 node cluster.
I dont know at this time, the data that need to be processed at this time for the Proof of
Concept. So - can you suggest something to me?
>>>>  
>>>> Regards,
>>>> Raj
>>> 
>>> 
>>> 
> 


Mime
View raw message