The newest Amazon HPC Compute Cluster Instance , how much it costs?
16 views
Skip to first unread message
Miha Ahronovitz
Jul 15, 2010, 3:09:23 AM7/15/10
to Cloud Computing
An 880 Amazon Compute Cluster instances ranked #146 costs over 8M
per year in Amazon fees alone assuming reserved instances and over
$12M per year without reservation, assuming no discounts otherthe
conditions from the their price list.
For more details see my blog
http://bit.ly/aRjnD2
All comments are welcome, including the ones pointing to errors and
oversights.
Miha
per year in Amazon fees alone assuming reserved instances and over
$12M per year without reservation, assuming no discounts otherthe
conditions from the their price list.
For more details see my blog
http://bit.ly/aRjnD2
All comments are welcome, including the ones pointing to errors and
oversights.
Miha
cloudsigma
Jul 19, 2010, 1:09:14 PM7/19/10
to Cloud Computing
Hi Miha,
I think it is worth pointing out that this latest announcement
represents Amazon expanding a bit their resource allocation range
although still keeping their resources bundled together tightly on a
proprietary platform which requires significant upfront investment to
get up and running. The following instance size details are directly
from Amazon's HPC page:
- 23 GB of memory
- 33.5 EC2 Compute Units (2 x Intel Xeon X5570, quad-core “Nehalem”
architecture)
- 1690 GB of instance storage
- 64-bit platform
- I/O Performance: Very High (10 Gigabit Ethernet)
API name: cc1.4xlarge
Comparing this to what already exists in our cloud (for example), I've
set out the current maximum instance size which is of course
unbundled:
- up to 32GB of RAM
- up to 20GHz CPU (with user definable cores up to 10 for multi-
threading)
- up to 1024 GB of persistent storage per drive with up to 4 drives
per instance
- 64bit or 32bit with fully user controlled operating system
- local storage on fast 2.5inch disks in RAID6 (with battery backed
hardware RAID cards)
We are by no means the only provider of such cloud computing in a more
flexible and higher performance way.
The biggest bottleneck for high performance computing is disk
performance which I think Amazon are relatively weak on. Adding 10Gbps
networking will help but not solve the constraint of having remote SAN
storage from the computing. Infiniband is much lower latency than
10Gbps networking if this really is your bottleneck.
Likewise the bundling of resources is highly constraining for most HPC
users who have pretty specific CPU/RAM/Storage ratios. The huge amount
of storage bundled with each instance compared with the CPU/RAM is a
great example of forced over-purchasing.
I think key performance points here are:
- getting the right resources ratio
- controlling the operating system to allow user-optimised
distributions for the particular task
- fast disks
In short:
Flexibility means increased applicability means higher performance
means more efficient computing means reduced costs
Best wishes,
Robert
--
Robert Jenkins
Co-founder
CloudSigma
http://www.cloudsigma.com/blog
http://www.twitter.com/CloudSigma
I think it is worth pointing out that this latest announcement
represents Amazon expanding a bit their resource allocation range
although still keeping their resources bundled together tightly on a
proprietary platform which requires significant upfront investment to
get up and running. The following instance size details are directly
from Amazon's HPC page:
- 23 GB of memory
- 33.5 EC2 Compute Units (2 x Intel Xeon X5570, quad-core “Nehalem”
architecture)
- 1690 GB of instance storage
- 64-bit platform
- I/O Performance: Very High (10 Gigabit Ethernet)
API name: cc1.4xlarge
Comparing this to what already exists in our cloud (for example), I've
set out the current maximum instance size which is of course
unbundled:
- up to 32GB of RAM
- up to 20GHz CPU (with user definable cores up to 10 for multi-
threading)
- up to 1024 GB of persistent storage per drive with up to 4 drives
per instance
- 64bit or 32bit with fully user controlled operating system
- local storage on fast 2.5inch disks in RAID6 (with battery backed
hardware RAID cards)
We are by no means the only provider of such cloud computing in a more
flexible and higher performance way.
The biggest bottleneck for high performance computing is disk
performance which I think Amazon are relatively weak on. Adding 10Gbps
networking will help but not solve the constraint of having remote SAN
storage from the computing. Infiniband is much lower latency than
10Gbps networking if this really is your bottleneck.
Likewise the bundling of resources is highly constraining for most HPC
users who have pretty specific CPU/RAM/Storage ratios. The huge amount
of storage bundled with each instance compared with the CPU/RAM is a
great example of forced over-purchasing.
I think key performance points here are:
- getting the right resources ratio
- controlling the operating system to allow user-optimised
distributions for the particular task
- fast disks
In short:
Flexibility means increased applicability means higher performance
means more efficient computing means reduced costs
Best wishes,
Robert
--
Robert Jenkins
Co-founder
CloudSigma
http://www.cloudsigma.com/blog
http://www.twitter.com/CloudSigma
Miha Ahronovitz
Jul 19, 2010, 2:15:49 PM7/19/10
to cloud-c...@googlegroups.com, Miha Ahronovitz
Robert, excellent post! Thanks for the details. What is the list price
for an on-demand for CloudSigma instance as described by you? As it is
unbundled, the price is not obvious and needs to be calculated. Also AWS
has a 1 year reserved instance and a 3 year reserved instance. How does
it work with CloudSigma?
for an on-demand for CloudSigma instance as described by you? As it is
unbundled, the price is not obvious and needs to be calculated. Also AWS
has a 1 year reserved instance and a 3 year reserved instance. How does
it work with CloudSigma?
Miha
On 7/19/2010 10:09 AM, cloudsigma wrote:
> Hi Miha,
>
> I think it is worth pointing out that this latest announcement
> represents Amazon expanding a bit their resource allocation range
> although still keeping their resources bundled together tightly on a
> proprietary platform which requires significant upfront investment to
> get up and running. The following instance size details are directly
> from Amazon's HPC page:
>
> - 23 GB of memory
> - 33.5 EC2 Compute Units (2 x Intel Xeon X5570, quad-core �Nehalem�
Miha Ahronovitz
Jul 19, 2010, 6:20:18 PM7/19/10
to Cloud Computing
Robert, regarding your disk performance observations:
> The biggest bottleneck for high performance computing is disk
> performance which I think Amazon are relatively weak on. Adding 10Gbps
> networking will help but not solve the constraint of having remote SAN
> storage from the computing. Infiniband is much lower latency than
> 10Gbps networking if this really is your bottleneck.
Here are some performed by Chris Dagdijian from Bio-Team
http://bit.ly/cmW8nT
Quote
" * Performance of the root/boot disk is way slower than any other
type of block based storage. This is to be expected as the boot disk
(even though it comes via an EBS-resident AMI) does not get the
benefit of paravirtualization acceleration. The take home message is
that the boot/root disk volume should not really be used for anything.
This also means that this blog post showing how to increase the size
of the local OS disk is useful only for playing around and not for
anything serious
* The performance of the ephemeral disks is better and striping
the two available drives together as a RAID0 volume has measurable
benefits across the board"
Miha
> CloudSigmahttp://www.cloudsigma.com/bloghttp://www.twitter.com/CloudSigma
> The biggest bottleneck for high performance computing is disk
> performance which I think Amazon are relatively weak on. Adding 10Gbps
> networking will help but not solve the constraint of having remote SAN
> storage from the computing. Infiniband is much lower latency than
> 10Gbps networking if this really is your bottleneck.
http://bit.ly/cmW8nT
Quote
" * Performance of the root/boot disk is way slower than any other
type of block based storage. This is to be expected as the boot disk
(even though it comes via an EBS-resident AMI) does not get the
benefit of paravirtualization acceleration. The take home message is
that the boot/root disk volume should not really be used for anything.
This also means that this blog post showing how to increase the size
of the local OS disk is useful only for playing around and not for
anything serious
* The performance of the ephemeral disks is better and striping
the two available drives together as a RAID0 volume has measurable
benefits across the board"
Miha
Ian Mills
Jul 20, 2010, 4:54:50 AM7/20/10
to cloud-c...@googlegroups.com
Latency and bandwidth are not the same thing. Latency is the time it takes to get to the 'other' end and depends on the speed of light, in practical terms a constant, and the length of the cable path. No matter how much bandwidth you buy the latency will stay the same. Coping with latency between disks and processors whilst maintaining integrity and performance at any distance is not easy.
--
Ian Mills
0755 394 6958
Service to the customer, the pursuit of excellence, respect for the individual.
--
~~~~~
UP 2010 Call For Proposals: New Cutting Edge Cloud Computing Conference http://www.up-con.com/content/call-proposals.
Official Cloud Slam websites - http://cloudslam10.com and http://cloudslam09.com
Posting guidelines: http://bit.ly/bL3u3v
Follow us on Twitter http://twitter.com/cloudcomp_group or @cloudcomp_group
Post Job/Resume at http://cloudjobs.net
Buy hundreds of conference sessions and panels on cloud computing on DVD at
http://www.amazon.com/gp/product/B002H07SEC, http://www.amazon.com/gp/product/B002H0IW1U or get instant access to downloadable versions at http://cloudslam09.com/content/registration-5.html and http://cloudslam10.com/content/registration
~~~~~
You received this message because you are subscribed to the Google Groups "Cloud Computing" group.
To post to this group, send email to cloud-c...@googlegroups.com
To unsubscribe from this group, send email to cloud-computi...@googlegroups.com
--
Ian Mills
0755 394 6958
Service to the customer, the pursuit of excellence, respect for the individual.
cloudsigma
Jul 21, 2010, 10:22:00 AM7/21/10
to Cloud Computing
Hi Miha,
In terms of CloudSigma we have burst and subscription pricing (the
equivalent of a reserved instance). For subscription we have 1 month,
3 months, 6 months and 1 year. We don't offer subscriptions for more
than 1 month as cloud computing in line with all computing is subject
to significant deflation due to Moore's law. We think it is
disingenuous to lock people into today's prices for 3 years during
which time the actual price of computing will more than halve.
For on-demand (what we call burst usage), the price varies depending
on our current cloud utilisation. We have floating prices based on a
fixed matrix with a T+5 pricing announcement made via RSS. This allows
users to set pricing tolerances and control when they do their
computing. You do this by scripting off the RSS feed subscription and
executing API calls based on the pricing levels. We will never shut
down a user's computing in the same way that spot instances work. This
puts the user in control. The T+5 delay allows a user to perform an
orderly shut-down of their computing should prices rise above the
tolerance level that they have set with their own monitoring scripts.
In terms of pricing I think the best way to look at this is how a real
user would do, i.e. in a cluster. I've priced up three prices for on-
demand and reserved usage. You will notice that for the 5 instance
cluster I have only included 1TB of storage. This is much more in line
with the real world usage we see from users of this type. They
generally don't have vast amounts of data in relation to their
computing needs. Yes they have a high volume of data but they
generally have a much higher volume of computing resource need to go
with it. I have reflected this in the example below. For on-demand
I've outlined the minimum and maximum as we have varying prices as set
out above.
5 instance cluster
100GHz CPU
100GB RAM
1TB storage
Reserved (1 month subscription) price: $6.49 per hour ($1.30 per hour
per instance)
Minimum on-demand (burst) price: $3.02 per hour ($0.60 per hour per
instance)
Maximum on-demand (burst) price: $9.92 per hour ($1.98 per hour per
instance)
One further important point, we have free unlimited incoming bandwidth
making it a zero cost proposition to migrate data to our cloud and to
keep updating it. This is important for computing as usually the raw
data i.e. the input is significantly larger then the process result
i.e. the output. Having free incoming data makes a significant cost
difference to this user profile. Amazon charge for incoming bandwidth
so this is an extra cost not directly included in a pure look at per
instance pricing.
Our outgoing data cost is just $0.0585 per GB. This compares with
Amazon's standard tariff of $0.15 per GB. Finally we don't throttle
users so when you do come to transfer data out after calculation, you
can do so at very high speed saving users a lot of time.
Kind regards,
Robert
--
> mij123.vcf
> < 1KViewDownload
In terms of CloudSigma we have burst and subscription pricing (the
equivalent of a reserved instance). For subscription we have 1 month,
3 months, 6 months and 1 year. We don't offer subscriptions for more
than 1 month as cloud computing in line with all computing is subject
to significant deflation due to Moore's law. We think it is
disingenuous to lock people into today's prices for 3 years during
which time the actual price of computing will more than halve.
For on-demand (what we call burst usage), the price varies depending
on our current cloud utilisation. We have floating prices based on a
fixed matrix with a T+5 pricing announcement made via RSS. This allows
users to set pricing tolerances and control when they do their
computing. You do this by scripting off the RSS feed subscription and
executing API calls based on the pricing levels. We will never shut
down a user's computing in the same way that spot instances work. This
puts the user in control. The T+5 delay allows a user to perform an
orderly shut-down of their computing should prices rise above the
tolerance level that they have set with their own monitoring scripts.
In terms of pricing I think the best way to look at this is how a real
user would do, i.e. in a cluster. I've priced up three prices for on-
demand and reserved usage. You will notice that for the 5 instance
cluster I have only included 1TB of storage. This is much more in line
with the real world usage we see from users of this type. They
generally don't have vast amounts of data in relation to their
computing needs. Yes they have a high volume of data but they
generally have a much higher volume of computing resource need to go
with it. I have reflected this in the example below. For on-demand
I've outlined the minimum and maximum as we have varying prices as set
out above.
5 instance cluster
100GHz CPU
100GB RAM
1TB storage
Reserved (1 month subscription) price: $6.49 per hour ($1.30 per hour
per instance)
Minimum on-demand (burst) price: $3.02 per hour ($0.60 per hour per
instance)
Maximum on-demand (burst) price: $9.92 per hour ($1.98 per hour per
instance)
One further important point, we have free unlimited incoming bandwidth
making it a zero cost proposition to migrate data to our cloud and to
keep updating it. This is important for computing as usually the raw
data i.e. the input is significantly larger then the process result
i.e. the output. Having free incoming data makes a significant cost
difference to this user profile. Amazon charge for incoming bandwidth
so this is an extra cost not directly included in a pure look at per
instance pricing.
Our outgoing data cost is just $0.0585 per GB. This compares with
Amazon's standard tariff of $0.15 per GB. Finally we don't throttle
users so when you do come to transfer data out after calculation, you
can do so at very high speed saving users a lot of time.
Kind regards,
Robert
--
> < 1KViewDownload
Reply all
Reply to author
Forward
0 new messages