Re: Cassandra hardware - balancing CPU/memory/iops/disk space

[Jesse McConnell]

let us know how the SSD's pan out, I am curious about that as well

cheers,
jesse

--
jesse mcconnell
jesse.mcconn...@gmail.com



On Tue, Mar 9, 2010 at 12:08, B. Todd Burruss  wrote:
> our dataset is too big to fit into cache, so we are hitting disk.  not a
> problem for normal operation, but when a node is restored, hinted handoff,
> load balanced, or if reads/write simply build up we see a problem.  the
> nodes can't seem to catch up.  this seems to be centered around drive seek
> time, not cassandra per se.
>
> to combat we are doing the following:
>
> - add more smaller drives per machine in RAID 0 to combat drive seek time.
> - scale horizontally - add more machines to cluster to spread the load
> - we also plan to try out SSDs as well.
>
>
> Jonathan Ellis wrote:
>>
>> Yes, but I would guess 90% of workloads are better served with
>> spending the extra money on more machines w/ cheap sata disks and lots
>> of ram.
>>
>> -Jonathan
>>
>> On Sun, Mar 7, 2010 at 1:00 PM, Boris Shulman  wrote:
>>
>>>
>>> Do you think having SAS disks will give better performance?
>>>
>>> On Sat, Mar 6, 2010 at 5:47 PM, Jonathan Ellis  wrote:
>>>

 I think http://wiki.apache.org/cassandra/CassandraHardware answers
 most of your questions.

 If possible, it's definitely useful to try out a small fraction of
 your anticipated workload against a test cluster, even a single node,
 before finalizing your production hardware purchase.

 On Sat, Mar 6, 2010 at 1:12 AM, Rosenberry, Eric
  wrote:

>
> I am looking for advice from others that are further along in deploying
> Cassandra in production environments than we are.  I want to know what
> you
> are finding your bottlenecks to be.  I would feel silly purchasing dual
> processor quad core 2.93ghz Nehalem machines with 192 gigs of RAM just
> to
> find out that the two local SATA disks kept all that CPU and RAM from
> being
> useful (clearly that example would be a dumb).
>
>
>
> I need to spec out hardware for an “optimal” Cassandra node (though our
> read/write characteristics are not yet fully defined so let’s go with
> an
> “average” configuration).
>
>
>
> My main concern is finding the right balance of:
>
> ·         Available CPU
>
> ·         RAM amount
>
> ·         RAM speed (think Nehalem architecture where memory comes in a
> few
> speeds, though I doubt this is much of a concern as it is mainly
> dictated by
> which processor you buy and how many slots you populate)
>
> ·         Total iops available (i.e. number of disks)
>
> ·         Total disk space available (depending on the ratio of
> iops/space
> deciding on SAS vs. SATA and various rotational speeds)
>
>
>
> My current thinking is 1U boxes with four 3.5 inch disks since that
> seems to
> be a readily available config.  One big question is should I go with a
> single processor Nehalem system to go with those four disks, or would
> two
> CPU’s be useful, and also, how much RAM is appropriate to match?  I am
> making the assumption that Cassandra nodes are going to be disk bound
> as
> they must do a random read to answer any given query (i.e. indexes in
> RAM,
> but all data lives on disk?).
>
>
>
> The other big decision is what type of hard disks others are finding to
> provide the optimal ratio of iops to available space?  SAS or SATA?
>  And
> what rotational speed?
>
>
>
> Let me throw out here an actual hardware config and feel free to tell
> me the
> error of my ways:
>
> ·         A SuperMicro SuperServer 6016T-NTRF configured as follows:
>
> o   2.26 ghz E5520 dual processor quad core hyperthreaded Nehalem
> architecture (this proc provides a lot of bang for the buck, faster
> procs
> get more expensive quickly)
>
> o   Qty 12, 4 gig 1066mhz DIMMS for a total of 48 gigs RAM (the 4 gig
> DIMMS
> seem to be the price sweet spot)
>
> o   Dual on board 1 gigabit NIC’s (perhaps one for client connections
> and
> the other for cluster communication?)
>
> o   Dual power supplies (I don’t want to lose half my cluster due to a
> failure on one power leg)
>
> o   4x 1TB SATA disks (this is a complete SWAG)
>
> o   No RAID controller (all just single individual disks presented to
> the
> OS) – Though is there any down side to using a RAID controller with
> RAID 0
> (perhaps one single disk for the log for sequential io’s, and 3x disks
> in a
> stripe for the random io’s)
>
> o   The on-board IPMI based OOB controller (so we can kick the boxes
> remotely if need be)
>
> ·
> http://www.supermicro.com/products/system/1U/6016/SYS-6016T-NTRF.cfm
>
>
>
> I can’t help but 

Re: Cassandra hardware - balancing CPU/memory/iops/disk space

[B. Todd Burruss]

our dataset is too big to fit into cache, so we are hitting disk.  not a 
problem for normal operation, but when a node is restored, hinted 
handoff, load balanced, or if reads/write simply build up we see a 
problem.  the nodes can't seem to catch up.  this seems to be centered 
around drive seek time, not cassandra per se.


to combat we are doing the following:

- add more smaller drives per machine in RAID 0 to combat drive seek time.
- scale horizontally - add more machines to cluster to spread the load
- we also plan to try out SSDs as well.


Jonathan Ellis wrote:

Yes, but I would guess 90% of workloads are better served with
spending the extra money on more machines w/ cheap sata disks and lots
of ram.

-Jonathan

On Sun, Mar 7, 2010 at 1:00 PM, Boris Shulman  wrote:
  

Do you think having SAS disks will give better performance?

On Sat, Mar 6, 2010 at 5:47 PM, Jonathan Ellis  wrote:


I think http://wiki.apache.org/cassandra/CassandraHardware answers
most of your questions.

If possible, it's definitely useful to try out a small fraction of
your anticipated workload against a test cluster, even a single node,
before finalizing your production hardware purchase.

On Sat, Mar 6, 2010 at 1:12 AM, Rosenberry, Eric
 wrote:
  

I am looking for advice from others that are further along in deploying
Cassandra in production environments than we are.  I want to know what you
are finding your bottlenecks to be.  I would feel silly purchasing dual
processor quad core 2.93ghz Nehalem machines with 192 gigs of RAM just to
find out that the two local SATA disks kept all that CPU and RAM from being
useful (clearly that example would be a dumb).



I need to spec out hardware for an “optimal” Cassandra node (though our
read/write characteristics are not yet fully defined so let’s go with an
“average” configuration).



My main concern is finding the right balance of:

· Available CPU

· RAM amount

· RAM speed (think Nehalem architecture where memory comes in a few
speeds, though I doubt this is much of a concern as it is mainly dictated by
which processor you buy and how many slots you populate)

· Total iops available (i.e. number of disks)

· Total disk space available (depending on the ratio of iops/space
deciding on SAS vs. SATA and various rotational speeds)



My current thinking is 1U boxes with four 3.5 inch disks since that seems to
be a readily available config.  One big question is should I go with a
single processor Nehalem system to go with those four disks, or would two
CPU’s be useful, and also, how much RAM is appropriate to match?  I am
making the assumption that Cassandra nodes are going to be disk bound as
they must do a random read to answer any given query (i.e. indexes in RAM,
but all data lives on disk?).



The other big decision is what type of hard disks others are finding to
provide the optimal ratio of iops to available space?  SAS or SATA?  And
what rotational speed?



Let me throw out here an actual hardware config and feel free to tell me the
error of my ways:

· A SuperMicro SuperServer 6016T-NTRF configured as follows:

o   2.26 ghz E5520 dual processor quad core hyperthreaded Nehalem
architecture (this proc provides a lot of bang for the buck, faster procs
get more expensive quickly)

o   Qty 12, 4 gig 1066mhz DIMMS for a total of 48 gigs RAM (the 4 gig DIMMS
seem to be the price sweet spot)

o   Dual on board 1 gigabit NIC’s (perhaps one for client connections and
the other for cluster communication?)

o   Dual power supplies (I don’t want to lose half my cluster due to a
failure on one power leg)

o   4x 1TB SATA disks (this is a complete SWAG)

o   No RAID controller (all just single individual disks presented to the
OS) – Though is there any down side to using a RAID controller with RAID 0
(perhaps one single disk for the log for sequential io’s, and 3x disks in a
stripe for the random io’s)

o   The on-board IPMI based OOB controller (so we can kick the boxes
remotely if need be)

·
http://www.supermicro.com/products/system/1U/6016/SYS-6016T-NTRF.cfm



I can’t help but think the above config has way too much RAM and CPU and not
enough iops capacity.  My understanding is that Cassandra does not cache
much in RAM though?



Any thoughts are appreciated.  Thanks.



-Eric

___
Eric Rosenberry
Sr. Infrastructure Architect | Chief Bit Plumber




iovation
111 SW Fifth Avenue
Suite 3200
Portland, OR 97204
www.iovation.com

The information contained in this email message may be privileged,
confidential and protected from disclosure. If you are not the intended
recipient, any dissemination, distribution or copying is strictly
prohibited. If you think that you have received this email message in error,
please notify the sender by reply email and delete the message and any
attachments.

Re: Cassandra hardware - balancing CPU/memory/iops/disk space

[Mark Robson]

On 7 March 2010 19:05, Jonathan Ellis  wrote:

> Yes, but I would guess 90% of workloads are better served with
> spending the extra money on more machines w/ cheap sata disks and lots
> of ram.
>

I'm not an expert, but I imagine that in many cases, capex is not the
limiting factor. In data centres, power appears to be the main limitations -
modern x86 servers use so much power that many older data centres cannot be
stocked at very high density (because the don't have enough cooling for that
much heat).

So the number of boxes you can put in is directly affected by the amount of
power they use. For this reason, the most efficient way is to find the ideal
combination of CPU, memory and disc to do what you want with the least
power.

I guess ram is pretty cheap on power - I mean, it uses hardly any.

Mark

Re: Cassandra hardware - balancing CPU/memory/iops/disk space

[Jonathan Ellis]

Yes, but I would guess 90% of workloads are better served with
spending the extra money on more machines w/ cheap sata disks and lots
of ram.

-Jonathan

On Sun, Mar 7, 2010 at 1:00 PM, Boris Shulman  wrote:
> Do you think having SAS disks will give better performance?
>
> On Sat, Mar 6, 2010 at 5:47 PM, Jonathan Ellis  wrote:
>> I think http://wiki.apache.org/cassandra/CassandraHardware answers
>> most of your questions.
>>
>> If possible, it's definitely useful to try out a small fraction of
>> your anticipated workload against a test cluster, even a single node,
>> before finalizing your production hardware purchase.
>>
>> On Sat, Mar 6, 2010 at 1:12 AM, Rosenberry, Eric
>>  wrote:
>>> I am looking for advice from others that are further along in deploying
>>> Cassandra in production environments than we are.  I want to know what you
>>> are finding your bottlenecks to be.  I would feel silly purchasing dual
>>> processor quad core 2.93ghz Nehalem machines with 192 gigs of RAM just to
>>> find out that the two local SATA disks kept all that CPU and RAM from being
>>> useful (clearly that example would be a dumb).
>>>
>>>
>>>
>>> I need to spec out hardware for an “optimal” Cassandra node (though our
>>> read/write characteristics are not yet fully defined so let’s go with an
>>> “average” configuration).
>>>
>>>
>>>
>>> My main concern is finding the right balance of:
>>>
>>> · Available CPU
>>>
>>> · RAM amount
>>>
>>> · RAM speed (think Nehalem architecture where memory comes in a few
>>> speeds, though I doubt this is much of a concern as it is mainly dictated by
>>> which processor you buy and how many slots you populate)
>>>
>>> · Total iops available (i.e. number of disks)
>>>
>>> · Total disk space available (depending on the ratio of iops/space
>>> deciding on SAS vs. SATA and various rotational speeds)
>>>
>>>
>>>
>>> My current thinking is 1U boxes with four 3.5 inch disks since that seems to
>>> be a readily available config.  One big question is should I go with a
>>> single processor Nehalem system to go with those four disks, or would two
>>> CPU’s be useful, and also, how much RAM is appropriate to match?  I am
>>> making the assumption that Cassandra nodes are going to be disk bound as
>>> they must do a random read to answer any given query (i.e. indexes in RAM,
>>> but all data lives on disk?).
>>>
>>>
>>>
>>> The other big decision is what type of hard disks others are finding to
>>> provide the optimal ratio of iops to available space?  SAS or SATA?  And
>>> what rotational speed?
>>>
>>>
>>>
>>> Let me throw out here an actual hardware config and feel free to tell me the
>>> error of my ways:
>>>
>>> · A SuperMicro SuperServer 6016T-NTRF configured as follows:
>>>
>>> o   2.26 ghz E5520 dual processor quad core hyperthreaded Nehalem
>>> architecture (this proc provides a lot of bang for the buck, faster procs
>>> get more expensive quickly)
>>>
>>> o   Qty 12, 4 gig 1066mhz DIMMS for a total of 48 gigs RAM (the 4 gig DIMMS
>>> seem to be the price sweet spot)
>>>
>>> o   Dual on board 1 gigabit NIC’s (perhaps one for client connections and
>>> the other for cluster communication?)
>>>
>>> o   Dual power supplies (I don’t want to lose half my cluster due to a
>>> failure on one power leg)
>>>
>>> o   4x 1TB SATA disks (this is a complete SWAG)
>>>
>>> o   No RAID controller (all just single individual disks presented to the
>>> OS) – Though is there any down side to using a RAID controller with RAID 0
>>> (perhaps one single disk for the log for sequential io’s, and 3x disks in a
>>> stripe for the random io’s)
>>>
>>> o   The on-board IPMI based OOB controller (so we can kick the boxes
>>> remotely if need be)
>>>
>>> ·
>>> http://www.supermicro.com/products/system/1U/6016/SYS-6016T-NTRF.cfm
>>>
>>>
>>>
>>> I can’t help but think the above config has way too much RAM and CPU and not
>>> enough iops capacity.  My understanding is that Cassandra does not cache
>>> much in RAM though?
>>>
>>>
>>>
>>> Any thoughts are appreciated.  Thanks.
>>>
>>>
>>>
>>> -Eric
>>>
>>> ___
>>> Eric Rosenberry
>>> Sr. Infrastructure Architect | Chief Bit Plumber
>>>
>>>
>>>
>>>
>>> iovation
>>> 111 SW Fifth Avenue
>>> Suite 3200
>>> Portland, OR 97204
>>> www.iovation.com
>>>
>>> The information contained in this email message may be privileged,
>>> confidential and protected from disclosure. If you are not the intended
>>> recipient, any dissemination, distribution or copying is strictly
>>> prohibited. If you think that you have received this email message in error,
>>> please notify the sender by reply email and delete the message and any
>>> attachments.
>>
>

Re: Cassandra hardware - balancing CPU/memory/iops/disk space

[Boris Shulman]

Do you think having SAS disks will give better performance?

On Sat, Mar 6, 2010 at 5:47 PM, Jonathan Ellis  wrote:
> I think http://wiki.apache.org/cassandra/CassandraHardware answers
> most of your questions.
>
> If possible, it's definitely useful to try out a small fraction of
> your anticipated workload against a test cluster, even a single node,
> before finalizing your production hardware purchase.
>
> On Sat, Mar 6, 2010 at 1:12 AM, Rosenberry, Eric
>  wrote:
>> I am looking for advice from others that are further along in deploying
>> Cassandra in production environments than we are.  I want to know what you
>> are finding your bottlenecks to be.  I would feel silly purchasing dual
>> processor quad core 2.93ghz Nehalem machines with 192 gigs of RAM just to
>> find out that the two local SATA disks kept all that CPU and RAM from being
>> useful (clearly that example would be a dumb).
>>
>>
>>
>> I need to spec out hardware for an “optimal” Cassandra node (though our
>> read/write characteristics are not yet fully defined so let’s go with an
>> “average” configuration).
>>
>>
>>
>> My main concern is finding the right balance of:
>>
>> · Available CPU
>>
>> · RAM amount
>>
>> · RAM speed (think Nehalem architecture where memory comes in a few
>> speeds, though I doubt this is much of a concern as it is mainly dictated by
>> which processor you buy and how many slots you populate)
>>
>> · Total iops available (i.e. number of disks)
>>
>> · Total disk space available (depending on the ratio of iops/space
>> deciding on SAS vs. SATA and various rotational speeds)
>>
>>
>>
>> My current thinking is 1U boxes with four 3.5 inch disks since that seems to
>> be a readily available config.  One big question is should I go with a
>> single processor Nehalem system to go with those four disks, or would two
>> CPU’s be useful, and also, how much RAM is appropriate to match?  I am
>> making the assumption that Cassandra nodes are going to be disk bound as
>> they must do a random read to answer any given query (i.e. indexes in RAM,
>> but all data lives on disk?).
>>
>>
>>
>> The other big decision is what type of hard disks others are finding to
>> provide the optimal ratio of iops to available space?  SAS or SATA?  And
>> what rotational speed?
>>
>>
>>
>> Let me throw out here an actual hardware config and feel free to tell me the
>> error of my ways:
>>
>> · A SuperMicro SuperServer 6016T-NTRF configured as follows:
>>
>> o   2.26 ghz E5520 dual processor quad core hyperthreaded Nehalem
>> architecture (this proc provides a lot of bang for the buck, faster procs
>> get more expensive quickly)
>>
>> o   Qty 12, 4 gig 1066mhz DIMMS for a total of 48 gigs RAM (the 4 gig DIMMS
>> seem to be the price sweet spot)
>>
>> o   Dual on board 1 gigabit NIC’s (perhaps one for client connections and
>> the other for cluster communication?)
>>
>> o   Dual power supplies (I don’t want to lose half my cluster due to a
>> failure on one power leg)
>>
>> o   4x 1TB SATA disks (this is a complete SWAG)
>>
>> o   No RAID controller (all just single individual disks presented to the
>> OS) – Though is there any down side to using a RAID controller with RAID 0
>> (perhaps one single disk for the log for sequential io’s, and 3x disks in a
>> stripe for the random io’s)
>>
>> o   The on-board IPMI based OOB controller (so we can kick the boxes
>> remotely if need be)
>>
>> ·
>> http://www.supermicro.com/products/system/1U/6016/SYS-6016T-NTRF.cfm
>>
>>
>>
>> I can’t help but think the above config has way too much RAM and CPU and not
>> enough iops capacity.  My understanding is that Cassandra does not cache
>> much in RAM though?
>>
>>
>>
>> Any thoughts are appreciated.  Thanks.
>>
>>
>>
>> -Eric
>>
>> ___
>> Eric Rosenberry
>> Sr. Infrastructure Architect | Chief Bit Plumber
>>
>>
>>
>>
>> iovation
>> 111 SW Fifth Avenue
>> Suite 3200
>> Portland, OR 97204
>> www.iovation.com
>>
>> The information contained in this email message may be privileged,
>> confidential and protected from disclosure. If you are not the intended
>> recipient, any dissemination, distribution or copying is strictly
>> prohibited. If you think that you have received this email message in error,
>> please notify the sender by reply email and delete the message and any
>> attachments.
>

Re: Cassandra hardware - balancing CPU/memory/iops/disk space

[Krishna Sankar]

Eric,
Couple of thoughts:
1. Hardware
Definitely dual quad core
12 X 4 DIMMS. This is the sweet spot for memory. I have many
machines with this config and some with the 12 X 2 configs
I haven¹t found the need for SATA and the higher price
Make sure you get good NICs
Are you using any virtualization layer ? I assume these are bare
metal with Ubuntu or RedHat.
2. Scaling
Naturally you should look at horizontal scaling than vertical.
An estimate of the application characteristics and data
properties would be helpful to get a first estimate
I think eventually you will end up with multiple boxes anyway,
so my philosophy has been to buy multiple optimal boxes
We are working on scaling characteristics (memory, network and
storage), unfortunately way too early to make any inferences
HTH.
Cheers

On 3/5/10 Fri Mar 5, 10, "Rosenberry, Eric" 
wrote:

> I am looking for advice from others that are further along in deploying
> Cassandra in production environments than we are.  I want to know what you are
> finding your bottlenecks to be.  I would feel silly purchasing dual processor
> quad core 2.93ghz Nehalem machines with 192 gigs of RAM just to find out that
> the two local SATA disks kept all that CPU and RAM from being useful (clearly
> that example would be a dumb).
>  
> I need to spec out hardware for an ³optimal² Cassandra node (though our
> read/write characteristics are not yet fully defined so let¹s go with an
> ³average² configuration).
>  
> My main concern is finding the right balance of:
> ·Available CPU
> 
> ·RAM amount
> 
> ·RAM speed (think Nehalem architecture where memory comes in a few
> speeds, though I doubt this is much of a concern as it is mainly dictated by
> which processor you buy and how many slots you populate)
> 
> ·Total iops available (i.e. number of disks)
> 
> ·Total disk space available (depending on the ratio of iops/space
> deciding on SAS vs. SATA and various rotational speeds)
> 
>  
> My current thinking is 1U boxes with four 3.5 inch disks since that seems to
> be a readily available config.  One big question is should I go with a single
> processor Nehalem system to go with those four disks, or would two CPU¹s be
> useful, and also, how much RAM is appropriate to match?  I am making the
> assumption that Cassandra nodes are going to be disk bound as they must do a
> random read to answer any given query (i.e. indexes in RAM, but all data lives
> on disk?).
>  
> The other big decision is what type of hard disks others are finding to
> provide the optimal ratio of iops to available space?  SAS or SATA?  And what
> rotational speed?
>  
> Let me throw out here an actual hardware config and feel free to tell me the
> error of my ways:
> ·A SuperMicro SuperServer 6016T-NTRF configured as follows:
> 
> o  2.26 ghz E5520 dual processor quad core hyperthreaded Nehalem architecture
> (this proc provides a lot of bang for the buck, faster procs get more
> expensive quickly)
> 
> o  Qty 12, 4 gig 1066mhz DIMMS for a total of 48 gigs RAM (the 4 gig DIMMS
> seem to be the price sweet spot)
> 
> o  Dual on board 1 gigabit NIC¹s (perhaps one for client connections and the
> other for cluster communication?)
> 
> o  Dual power supplies (I don¹t want to lose half my cluster due to a failure
> on one power leg)
> 
> o  4x 1TB SATA disks (this is a complete SWAG)
> 
> o  No RAID controller (all just single individual disks presented to the OS) ­
> Though is there any down side to using a RAID controller with RAID 0 (perhaps
> one single disk for the log for sequential io¹s, and 3x disks in a stripe for
> the random io¹s)
> 
> o  The on-board IPMI based OOB controller (so we can kick the boxes remotely
> if need be)
> 
> ·http://www.supermicro.com/products/system/1U/6016/SYS-6016T-NTRF.cfm
> 
>  
> I can¹t help but think the above config has way too much RAM and CPU and not
> enough iops capacity.  My understanding is that Cassandra does not cache much
> in RAM though?
>  
> Any thoughts are appreciated.  Thanks.
>  
> -Eric
> ___
> Eric Rosenberry
> Sr. Infrastructure Architect | Chief Bit Plumber
>  
>  
> iovation
> 111 SW Fifth Avenue
> Suite 3200
> Portland, OR 97204
> www.iovation.com 
>  
> The information contained in this email message may be privileged,
> confidential and protected from disclosure. If you are not the intended
> recipient, any dissemination, distribution or copying is strictly prohibited.
> If you think that you have received this email message in error, please notify
> the sender by reply email and delete the message and any attachments.
> 

Re: Cassandra hardware - balancing CPU/memory/iops/disk space

[Jonathan Ellis]

I think http://wiki.apache.org/cassandra/CassandraHardware answers
most of your questions.

If possible, it's definitely useful to try out a small fraction of
your anticipated workload against a test cluster, even a single node,
before finalizing your production hardware purchase.

On Sat, Mar 6, 2010 at 1:12 AM, Rosenberry, Eric
 wrote:
> I am looking for advice from others that are further along in deploying
> Cassandra in production environments than we are.  I want to know what you
> are finding your bottlenecks to be.  I would feel silly purchasing dual
> processor quad core 2.93ghz Nehalem machines with 192 gigs of RAM just to
> find out that the two local SATA disks kept all that CPU and RAM from being
> useful (clearly that example would be a dumb).
>
>
>
> I need to spec out hardware for an “optimal” Cassandra node (though our
> read/write characteristics are not yet fully defined so let’s go with an
> “average” configuration).
>
>
>
> My main concern is finding the right balance of:
>
> · Available CPU
>
> · RAM amount
>
> · RAM speed (think Nehalem architecture where memory comes in a few
> speeds, though I doubt this is much of a concern as it is mainly dictated by
> which processor you buy and how many slots you populate)
>
> · Total iops available (i.e. number of disks)
>
> · Total disk space available (depending on the ratio of iops/space
> deciding on SAS vs. SATA and various rotational speeds)
>
>
>
> My current thinking is 1U boxes with four 3.5 inch disks since that seems to
> be a readily available config.  One big question is should I go with a
> single processor Nehalem system to go with those four disks, or would two
> CPU’s be useful, and also, how much RAM is appropriate to match?  I am
> making the assumption that Cassandra nodes are going to be disk bound as
> they must do a random read to answer any given query (i.e. indexes in RAM,
> but all data lives on disk?).
>
>
>
> The other big decision is what type of hard disks others are finding to
> provide the optimal ratio of iops to available space?  SAS or SATA?  And
> what rotational speed?
>
>
>
> Let me throw out here an actual hardware config and feel free to tell me the
> error of my ways:
>
> · A SuperMicro SuperServer 6016T-NTRF configured as follows:
>
> o   2.26 ghz E5520 dual processor quad core hyperthreaded Nehalem
> architecture (this proc provides a lot of bang for the buck, faster procs
> get more expensive quickly)
>
> o   Qty 12, 4 gig 1066mhz DIMMS for a total of 48 gigs RAM (the 4 gig DIMMS
> seem to be the price sweet spot)
>
> o   Dual on board 1 gigabit NIC’s (perhaps one for client connections and
> the other for cluster communication?)
>
> o   Dual power supplies (I don’t want to lose half my cluster due to a
> failure on one power leg)
>
> o   4x 1TB SATA disks (this is a complete SWAG)
>
> o   No RAID controller (all just single individual disks presented to the
> OS) – Though is there any down side to using a RAID controller with RAID 0
> (perhaps one single disk for the log for sequential io’s, and 3x disks in a
> stripe for the random io’s)
>
> o   The on-board IPMI based OOB controller (so we can kick the boxes
> remotely if need be)
>
> ·
> http://www.supermicro.com/products/system/1U/6016/SYS-6016T-NTRF.cfm
>
>
>
> I can’t help but think the above config has way too much RAM and CPU and not
> enough iops capacity.  My understanding is that Cassandra does not cache
> much in RAM though?
>
>
>
> Any thoughts are appreciated.  Thanks.
>
>
>
> -Eric
>
> ___
> Eric Rosenberry
> Sr. Infrastructure Architect | Chief Bit Plumber
>
>
>
>
> iovation
> 111 SW Fifth Avenue
> Suite 3200
> Portland, OR 97204
> www.iovation.com
>
> The information contained in this email message may be privileged,
> confidential and protected from disclosure. If you are not the intended
> recipient, any dissemination, distribution or copying is strictly
> prohibited. If you think that you have received this email message in error,
> please notify the sender by reply email and delete the message and any
> attachments.

Cassandra hardware - balancing CPU/memory/iops/disk space

[Rosenberry, Eric]

I am looking for advice from others that are further along in deploying 
Cassandra in production environments than we are.  I want to know what you are 
finding your bottlenecks to be.  I would feel silly purchasing dual processor 
quad core 2.93ghz Nehalem machines with 192 gigs of RAM just to find out that 
the two local SATA disks kept all that CPU and RAM from being useful (clearly 
that example would be a dumb).

I need to spec out hardware for an "optimal" Cassandra node (though our 
read/write characteristics are not yet fully defined so let's go with an 
"average" configuration).

My main concern is finding the right balance of:

* Available CPU

* RAM amount

* RAM speed (think Nehalem architecture where memory comes in a few 
speeds, though I doubt this is much of a concern as it is mainly dictated by 
which processor you buy and how many slots you populate)

* Total iops available (i.e. number of disks)

* Total disk space available (depending on the ratio of iops/space 
deciding on SAS vs. SATA and various rotational speeds)

My current thinking is 1U boxes with four 3.5 inch disks since that seems to be 
a readily available config.  One big question is should I go with a single 
processor Nehalem system to go with those four disks, or would two CPU's be 
useful, and also, how much RAM is appropriate to match?  I am making the 
assumption that Cassandra nodes are going to be disk bound as they must do a 
random read to answer any given query (i.e. indexes in RAM, but all data lives 
on disk?).

The other big decision is what type of hard disks others are finding to provide 
the optimal ratio of iops to available space?  SAS or SATA?  And what 
rotational speed?

Let me throw out here an actual hardware config and feel free to tell me the 
error of my ways:

* A SuperMicro SuperServer 6016T-NTRF configured as follows:

o   2.26 ghz E5520 dual processor quad core hyperthreaded Nehalem architecture 
(this proc provides a lot of bang for the buck, faster procs get more expensive 
quickly)

o   Qty 12, 4 gig 1066mhz DIMMS for a total of 48 gigs RAM (the 4 gig DIMMS 
seem to be the price sweet spot)

o   Dual on board 1 gigabit NIC's (perhaps one for client connections and the 
other for cluster communication?)

o   Dual power supplies (I don't want to lose half my cluster due to a failure 
on one power leg)

o   4x 1TB SATA disks (this is a complete SWAG)

o   No RAID controller (all just single individual disks presented to the OS) - 
Though is there any down side to using a RAID controller with RAID 0 (perhaps 
one single disk for the log for sequential io's, and 3x disks in a stripe for 
the random io's)

o   The on-board IPMI based OOB controller (so we can kick the boxes remotely 
if need be)

* http://www.supermicro.com/products/system/1U/6016/SYS-6016T-NTRF.cfm

I can't help but think the above config has way too much RAM and CPU and not 
enough iops capacity.  My understanding is that Cassandra does not cache much 
in RAM though?

Any thoughts are appreciated.  Thanks.

-Eric
___
Eric Rosenberry
Sr. Infrastructure Architect | Chief Bit Plumber


iovation
111 SW Fifth Avenue
Suite 3200
Portland, OR 97204
www.iovation.com

The information contained in this email message may be privileged, confidential 
and protected from disclosure. If you are not the intended recipient, any 
dissemination, distribution or copying is strictly prohibited. If you think 
that you have received this email message in error, please notify the sender by 
reply email and delete the message and any attachments.