Dear friends,

Please suggest a standard hardware configuration for hbase cluster which  is
going to be used to pull and store a lot of data.

--
Thanks,
Shah

http://www.cloudera.com/blog/2010/03/clouderas-support-team-shares-some-basic-hardware-recommendations/

"4 1TB hard disks in a JBOD (Just a Bunch Of Disks) configuration
2 quad core CPUs, running at least 2-2.5GHz
16-24GBs of RAM (24-32GBs if you’re considering HBase)
Gigabit Ethernet"

HTH,
Tim

On Tue, Jun 7, 2011 at 10:11 AM, Shahnawaz Saifi <[EMAIL PROTECTED]> wrote:
> Dear friends,
>
> Please suggest a standard hardware configuration for hbase cluster which  is
> going to be used to pull and store a lot of data.
>
> --
> Thanks,
> Shah
>

Lots of people are moving towards more spindles per box to increase IOP/s

This is particular important for cases where the working set gets pushed out
of memory.

On Tue, Jun 7, 2011 at 1:32 AM, Tim Robertson <[EMAIL PROTECTED]>wrote:

>
> http://www.cloudera.com/blog/2010/03/clouderas-support-team-shares-some-basic-hardware-recommendations/
>
> "4 1TB hard disks in a JBOD (Just a Bunch Of Disks) configuration
> 2 quad core CPUs, running at least 2-2.5GHz
> 16-24GBs of RAM (24-32GBs if you’re considering HBase)
> Gigabit Ethernet"
>
> HTH,
> Tim
>
>
>
> On Tue, Jun 7, 2011 at 10:11 AM, Shahnawaz Saifi <[EMAIL PROTECTED]>
> wrote:
> > Dear friends,
> >
> > Please suggest a standard hardware configuration for hbase cluster which
>  is
> > going to be used to pull and store a lot of data.
> >
> > --
> > Thanks,
> > Shah
> >
>

> From: Ted Dunning <[EMAIL PROTECTED]>
> Lots of people are moving towards more spindles per box to
> increase IOP/s
>
> This is particular important for cases where the working
> set gets pushed out of memory.

Indeed.

Our spec is more like 12x 500 GB SATA disks, to push IOPS and more evenly balance CPUs (fast dual quad core) with resident data per box. 10 of those disks are JBOD disks for the DataNodes. Two are RAID-1 system disk so dumb errors on the system disk doesn't bring down all volumes.

  - Andy

Ensure enough networking bandwidth to match your drive-bandwidth, otherwise
your compaction rates are going to be abysmal. 10 GigE ports are expensive,
so consider  2 x 1GigE per box (or even 4 x 1GigE if you can get that many
on-board NICs).
On Wed, Jun 8, 2011 at 8:49 AM, Andrew Purtell <[EMAIL PROTECTED]> wrote:

> > From: Ted Dunning <[EMAIL PROTECTED]>
> > Lots of people are moving towards more spindles per box to
> > increase IOP/s
> >
> > This is particular important for cases where the working
> > set gets pushed out of memory.
>
> Indeed.
>
> Our spec is more like 12x 500 GB SATA disks, to push IOPS and more evenly
> balance CPUs (fast dual quad core) with resident data per box. 10 of those
> disks are JBOD disks for the DataNodes. Two are RAID-1 system disk so dumb
> errors on the system disk doesn't bring down all volumes.
>
>  - Andy
>
>

Expensive is relative and with the latest intel hardware release you're starting to see 10gbe on the motherboard.

Not to mention you don't get a linear boost w port bonding.

You have to be careful on hardware recommendations because there are pricing sweet spots and technology changes.

Sent from a remote device. Please excuse any typos...

Mike Segel

On Jun 9, 2011, at 12:14 PM, "M. C. Srivas" <[EMAIL PROTECTED]> wrote:

> Ensure enough networking bandwidth to match your drive-bandwidth, otherwise
> your compaction rates are going to be abysmal. 10 GigE ports are expensive,
> so consider  2 x 1GigE per box (or even 4 x 1GigE if you can get that many
> on-board NICs).
>
>
> On Wed, Jun 8, 2011 at 8:49 AM, Andrew Purtell <[EMAIL PROTECTED]> wrote:
>
>>> From: Ted Dunning <[EMAIL PROTECTED]>
>>> Lots of people are moving towards more spindles per box to
>>> increase IOP/s
>>>
>>> This is particular important for cases where the working
>>> set gets pushed out of memory.
>>
>> Indeed.
>>
>> Our spec is more like 12x 500 GB SATA disks, to push IOPS and more evenly
>> balance CPUs (fast dual quad core) with resident data per box. 10 of those
>> disks are JBOD disks for the DataNodes. Two are RAID-1 system disk so dumb
>> errors on the system disk doesn't bring down all volumes.
>>
>> - Andy
>>
>>

On Fri, Jun 10, 2011 at 3:58 AM, Michel Segel <[EMAIL PROTECTED]>wrote:

> Not to mention you don't get a linear boost w port bonding.
>

Well, you don't get linear boost with switch level bonding.  You can get it,
however.
> You have to be careful on hardware recommendations because there are
> pricing sweet spots and technology changes.
>

This is absolutely true.  These sweet spots change on a nearly monthly basis
as well.  Frequent re-evaluation is critical.

And even that recommendation isn't right. ;-)

I think Sandy Bridge and SolarFlare are changing some of the design considerations.

> Date: Tue, 7 Jun 2011 10:32:58 +0200
> Subject: Re: Hbase Hardware requirement
> From: [EMAIL PROTECTED]
> To: [EMAIL PROTECTED]
>
> http://www.cloudera.com/blog/2010/03/clouderas-support-team-shares-some-basic-hardware-recommendations/
>
> "4 1TB hard disks in a JBOD (Just a Bunch Of Disks) configuration
> 2 quad core CPUs, running at least 2-2.5GHz
> 16-24GBs of RAM (24-32GBs if you’re considering HBase)
> Gigabit Ethernet"
>
> HTH,
> Tim
>
>
>
> On Tue, Jun 7, 2011 at 10:11 AM, Shahnawaz Saifi <[EMAIL PROTECTED]> wrote:
> > Dear friends,
> >
> > Please suggest a standard hardware configuration for hbase cluster which  is
> > going to be used to pull and store a lot of data.
> >
> > --
> > Thanks,
> > Shah
> >
     

Depends on the load.  We have huge cluster running, 4 x 2 TB disks,
Core 2 Duo 2.5 Ghz, 8 GB RAM, with 60 nodes, using it mostly for
binary cold storage of photos, with very low access rates, and
moderate write rates.

Second cluster, is Core i7 Quad (hyperthreaded) 3.0Ghz , with 16GB
RAM, 4x2TB drives 20 nodes.  Powering yfrog.com metadata, lots of
writes and lots of reads, user facing.  The cost per node is about
$1200.00 (commodity desktop mobo).

-Jack

On Tue, Jun 7, 2011 at 4:13 AM, Michael Segel <[EMAIL PROTECTED]> wrote:
>
> And even that recommendation isn't right. ;-)
>
> I think Sandy Bridge and SolarFlare are changing some of the design considerations.
>
>> Date: Tue, 7 Jun 2011 10:32:58 +0200
>> Subject: Re: Hbase Hardware requirement
>> From: [EMAIL PROTECTED]
>> To: [EMAIL PROTECTED]
>>
>> http://www.cloudera.com/blog/2010/03/clouderas-support-team-shares-some-basic-hardware-recommendations/
>>
>> "4 1TB hard disks in a JBOD (Just a Bunch Of Disks) configuration
>> 2 quad core CPUs, running at least 2-2.5GHz
>> 16-24GBs of RAM (24-32GBs if you’re considering HBase)
>> Gigabit Ethernet"
>>
>> HTH,
>> Tim
>>
>>
>>
>> On Tue, Jun 7, 2011 at 10:11 AM, Shahnawaz Saifi <[EMAIL PROTECTED]> wrote:
>> > Dear friends,
>> >
>> > Please suggest a standard hardware configuration for hbase cluster which  is
>> > going to be used to pull and store a lot of data.
>> >
>> > --
>> > Thanks,
>> > Shah
>> >
>

In my experience the network is the bottleneck in a lot of processing jobs where you read a row from HBase, do some processing, and write back some additional data.  I would suggest 10 GigE.
Dave
-----Original Message-----
From: Shahnawaz Saifi [mailto:[EMAIL PROTECTED]]
Sent: Tuesday, June 07, 2011 1:11 AM
To: [EMAIL PROTECTED]
Subject: Hbase Hardware requirement

Dear friends,

Please suggest a standard hardware configuration for hbase cluster which  is
going to be used to pull and store a lot of data.

--
Thanks,
Shah