You are on page 1of 10

1 SeIectIng a Ircessr Ir SI Server 2012

Saturday 07 june 2014 09.22.42 ST http.//sqIperIrmance.cm/2013/01/system-cnIIguratIn/seIectIng-...


Posted by Clenn Berry on anuary 9 , 2013 Add comments
5eIectinga Processor or 5QL5erver 2012

Since Microsoft revamped the licensing model for SQL Server 2012, it is
especially important to do some thoughtful analysis before you decide
exactly which processor to use for a database server that will be running SQL
Server 2012. The move to core-based licensing for SQL Server 2012 Enterprise
Edition means that a careless decision about precisely what processor you
will be running on can cost both a great deal of money and a great deal of
performance and scalability. The same issue exists (to a lesser extent) with
SQL Server 2012 Standard Edition.
The difference in SQL Server 2012 licensing costs between a good processor
choice and a bad processor choice can more than pay for your hardware and
your storage subsystem in many cases. Civen this, how do you go about
making an optimal processor choice for SQL Server 2012 Enterprise Edition7
Normally, the first step would be to decide whether you wanted to use an
ntel Xeon processor or an AMD Opteron processor (SQL Server 2012 does not
support the ntel tanium processor family).
Unfortunately, it is very difficult to make a good technical or business case to
select an AMD Opteron processor for use with SQL Server 2012 Enterprise
Edition. Since the introduction of the ntel Nehalem architecture (ntel Xeon
3500, 5500, 6500 and 7500 series) in 2008-2010, AMD has simply not been able
to compete with ntel when it comes to single-threaded processor
performance. AMD does not have the financial or technical resources to
compete with ntel in terms of performance or power efficiency. With each
new processor family release from ntel, AMD has been falling further behind.
This is not a good thing for the .T. industry, since a lack of viable
competition from AMD will inevitably slow down the pace of innovation from
ntel .
Since an ntel Xeon processor seems to be the only viable choice for SQL
Server 2012, the next step is deciding which one of the many available Xeon
families and models would be the best choice for your intended SQL Server
2012 workload. ntel has different Xeon product families for different socket
count servers. For single-socket workstations and servers they have the ntel
Xeon E3 family. For two-socket workstations and servers they have the ntel
Xeon E5 family. Finally, for two-socket, four-socket, and eight-socket (or
more) servers, they have the ntel Xeon E7 family.
Since this article is discussing SQL Server 2012 Enterprise Edition, we will
ignore the single-socket ntel Xeon E3 family, since ntel Xeon E3 processors
Search
5earch 5QLPerormance. com
Aaron Bertrand
Ben|amin Nevarez
Erin Stellato
Clenn Berry
ason Hall
oe Sack
onathan Kehayias
Kevin Kline
Paul Randal
Paul White
Rick Pittser
Authors
Database Design (7)
Error Handling (1)
Extended Events (8)
nstallation ( 6)
OSubsystem (18)
Service Broker ( 1)
SQL ndexes ( 28)
SQL ob (1)
SQL Lock/Block/Deadlock (3)
SQL Maintenance (3)
SQL Memory (6)
SQL Optimizer (21)
SQL Performance ( 112)
SQL Plan ( 45)
SQL Statistics (7)
SQL Trace (5)
SystemConfiguration (34)
T-SQL Queries ( 73)
Transaction Log (11)
Waits ( 8)
Categories
une 2014 (1)
May 2014 (6)
April 2014 ( 6)
March 2014 (7)
Archives
About Contact RSS Feed
Home Q & A
2 SeIectIng a Ircessr Ir SI Server 2012
Saturday 07 june 2014 09.22.42 ST http.//sqIperIrmance.cm/2013/01/system-cnIIguratIn/seIectIng-...
are limited to using 32CB of DDR3 RAM. There are some niche scenarios where
it might make very good sense to use an ntel Xeon E3-1290v2 processor in a
single-socket server with 32CB of RAM in combination with SQL Server 2012
Enterprise Edition. Perhaps you have a relatively small database where you
need the absolute fastest single-threaded performance and you also need
specific Enterprise Edition features such as SQL Server AlwaysOn Availability
Croups.
The more common choice is between a two-socket server and a four-socket
(or more) server. You are going to have to assess your workload size and
intensity and decide whether it can run on a smaller, but faster two-socket
server, or whether you will have to make the |ump to a larger, slower four-
socket or larger server. This decision is directly affected by your required total
CPU capacity, total physical RAM capacity, and your total required /O
capacity (which is related to the number and type of PC -E expansion slots in
the server).
One common misconception is that bigger ntel-based servers (in terms of
socket counts) are faster servers. This is simply not true, for a number of
reasons. The sales volume and market share of two- socket servers is much
higher than it is for four-socket and larger servers. There is also less
engineering and validation work required for two-socket capable ntel
processors compared to four-socket capable ntel processors. Because of
these factors, ntel releases new processor architectures more frequently and
earlier for lower socket count servers. Currently, ntel' s single-socket E3 family
is using the 22nm vy Bridge and the two-socket E5 family is using the 32nm
Sandy Bridge-EP, while ntel E7 family is using the older 32nm Westmere-EX
microarchitecture.
Another reason is that you do not get linear scaling as you increase your
socket count, even with Non-uniform memory access (NUMA) architecture
processors, which scale much better than the older symmetrical
multiprocessing (SMP) architecture. This means that a four-socket server will
not have twice the processor performance or capacity as a two-socket server
with the same model processor.
This can be confirmed by comparing the TPC-E OLTP benchmark results of
two-socket systems with ntel Xeon E7-2870 processors to four-socket systems
with ntel Xeon E7-4870 processors to eight-socket systems with ntel Xeon
E7-8870 processors. Even though these are essentially the same processor
with the same individual performance characteristics, the TPC-E benchmark
score does not double as you double the socket count, as you can see in
Table 1.
Processor
5ocket
Count
TPC-E
5core
TotaI Core
Count
TPC-E
5core/Core
Xeon E7-2870 2 1560.70 20 78.04
Xeon E7-4870 4 2862.61 40 71.57
Xeon E7-8870 8 4614.22 80 57.68
Table 1: Comparison of TPC-E Scores as Socket Count ncreases
When think about comparing single-socket to two-socket, to four and eight-
socket processors, like to use a car and truck analogy. A single-socket server
February 2014 ( 5)
anuary 2014 (9)
December 2013 ( 5)
November 2013 (5)
October 2013 (7)
September 2013 (6)
August 2013 (6)
uly 2013 (7)
une 2013 (6)
May 2013 (5)
April 2013 ( 8)
March 2013 (7)
February 2013 ( 9)
anuary 2013 (6)
December 2012 ( 4)
November 2012 (7)
October 2012 (7)
September 2012 (4)
August 2012 (9)
uly 2012 (2)
3 SeIectIng a Ircessr Ir SI Server 2012
Saturday 07 june 2014 09.22.42 ST http.//sqIperIrmance.cm/2013/01/system-cnIIguratIn/seIectIng-...
Tagged with: CPUs, hardware, processors

19 Responses to "Sel ecti ng a Processor for SQL Serv er 2012"
1. Rat eCont rol says:
anuary 9, 2013 at 4:29 PM
Creat article. Thanks for putting it together!
Reply
Cl enn Berry says:
anuary 9, 2013 at 6:01 PM
Clad you liked it.



is like a Formula-1 race car, being extremely fast but having very little cargo
capacity. A two-socket server is like a Tesla Model S, being very fast and
having pretty decent cargo capacity. A four-socket server is like a large SUv,
being slower but having more cargo capacity than a Tesla Model S. Finally, an
eight-socket server is like a Mack truck, able to haul a huge load at a much
slower rate than an SUv.
Processor
5ocket
Count
TPC-E
5core
TotaI Core
Count
TPC-E
5core/Core
Xeon E5-2690 2 1881.76 16 117.61
Xeon E5-4650 4 2651.27 32 82.85
Table 2: Comparison of TPC-E Scores for Two Xeon E5 Processor Models
Comparing Table 1 to Table 2, we can see that the ntel Xeon E5 family does
quite a bit better on TPC-E than the ntel Xeon E7 family does, which is no
surprise, since we are comparing the newer Sandy Bridge-EP to the older
Westmere-EX microarchitecture. From a performance perspective, the two-
socket Xeon E5-2690 does much better than the two-socket Xeon E7-2870. n
my opinion, you really should not be using the two-socket Xeon E7-2870 for
SQL Server 2012 because of its lower single-threaded performance and higher
physical core counts (which means a higher SQL Server 2012 licensing cost).
Currently, my favorite ntel server processor is the ntel Xeon E5-2690. t will
give you excellent single-threaded performance and relatively affordable SQL
Server 2012 licensing costs. f you need to step up to a four-socket server,
then would choose an ntel Xeon E5-4650 processor instead of using an ntel
Xeon E7-4870 processor, since you will get better single-threaded performance
and lower SQL Server 2012 license costs. Using TPC-E benchmark scores is an
excellent way to compare the performance and SQL Server 2012 license
efficiency of different processor families.
4 SeIectIng a Ircessr Ir SI Server 2012
Saturday 07 june 2014 09.22.42 ST http.//sqIperIrmance.cm/2013/01/system-cnIIguratIn/seIectIng-...
Reply
2. Creg Li nwood says:
anuary 9, 2013 at 9:20 PM
Hi Clenn
Cood article, however the choice of CPU alsoaffects long termpower consumption
costs which 'd be interested in hearing your thoughts on. agree that ntel provides
better rawpower than AMD per thread & the cost implications under SQL 2012' s
licensing model magnify the difference in per core performance between AMD & ntel.
However , over the 3 > 4 years lifecycle (sometimes more) of a SQL Server solution, the
total power consumption costs are also an important factor in the overall costs. These
costs are usually buried in the overall data centre costs but power costs are growing
faster than the cost of SQL Server licensing in many parts of the world so this factor is
also growing in importance.
Have you attempted any comparison of the total cost / performance difference
between AMD & ntel7
Regards,
Creg Linwood
Reply
3. Cl enn Berry says:
anuary 9, 2013 at 9:35 PM
agree that power consumption and power costs can be a significant long-term
expense. From what testing have seen, the ntel Xeon E5 Family has much better
performance/watt than the AMD Opteron 6200 Family . The AMD Opteron 6300 Family
is a little better than the 6200 Family in performance/watt, but still not as good as
ntel.
http://www.anandtech.com/show/5553/the-xeon-e52600- dual-sandybridge- for-
servers/7
Another factor to consider is that a higher performance processor can complete its
workload more quickly and be able toreduce its power consumption more often
(depending on your power management settings), so that will often have a ma|or
effect on overall power consumption. AnandTech has done some testing toconfirm
this.
Reply
4. kev i ncl osson says:
anuary 9, 2013 at 10:12 PM
Crystal -clear NUMA effect on the WSM- EX ( aka E7) kit. We see 78, 72, 58 TPC-E/core at
2,4 and 8 sockets. With 2 and 4 sockets there is point-to-point memory connectivity
since WSM-EX has 3 QP links. umping to8 requires glue (ala BM eXA or HP PRMA)
else one ends up with 2-hop memory as is the case with the Sun x4800 (aka Sun
Server X2-8). What brand server was that E7-8870 TPC-E result obtained on7
As for Sandy Bridge and per -core licensing actually would rather see folks consider
the E5-2643 which is *exaclty * an E5-2690 chopped in half so it is 4c/8t. f someone is
replatforming from, say, Harpertown Xeon ( also4c but noSMT) they can easily expect
4.5-5x performance improvement when |umping to E5-2643- if SPEC is any indicator
(and find it usually is). So folks, do yourself a favor and at least test a 2S server fitted
with E5-2643. knowHP Proliant and Cisco UCS (rack-mounted) are offering this ntel
SKU as an option.
5 SeIectIng a Ircessr Ir SI Server 2012
Saturday 07 june 2014 09.22.42 ST http.//sqIperIrmance.cm/2013/01/system-cnIIguratIn/seIectIng-...
Nice article, Clenn
kevinclosson.wordpress.com
Reply
Cl enn Berry says:
anuary 9, 2013 at 11:58 PM
Thanks, Kevin! The E7-8870 TPC- E result was an NEC Express5800/A1080a-
E. have recommended the E5-2643 topeople to save on SQL Server 2012
licensing costs quite a few times in the past. t actually has a higher base
clock speed (3.3CHz) than the E5-2690 ( 2.9CHz), although it has a lower
Turbo Boost speed (3.5CHz for the E5-2643) vs. (3.8CHz for the E5-2690).
Either one of those can easily replace a four-socket server from3-4 years
ago in most cases.
Reply
kev i ncl osson says:
anuary 10, 2013 at 4:52 PM
Hi Clenn,
Actually, both the 2690 and 2643 run at 3.4CHz when only 4
cores are active. The 2643 is really a 2690 chopped in half (with
on-die logic) . The 2690 core frequency drops to 3.3CHz once 6
cores are active. The 2643 drops to3.4 once 3 cores are active.
The difference in these core frequencies matter little for load/
store workloads like databases. the amped up frequencies
don' t help stalls :-) Besides, that's only 3% difference anyway .
The bigger point- that you make well-is the fact that these SQL
Server customers should choose their SKUs when re-
platforming fromvintage 4yr old systems. On a per core basis
E5 top-bin SKUs are about 4-5x more powerful (e.g., Xeon
5400 compared toE5-2643) .
Reply
5. Creg Li nwood says:
anuary 9, 2013 at 10:34 PM
t's a pity that so fewTPC- E results are reporting Watts/tpsE as this would be more
relevant toSQL and save us fromhaving tointerpret what more generic tests such as
AnandTech mean with a SQL workload which usually has heavy /Ocharacteristics
&CPU idle time.
The fewTPC-E results which have been published with Watts/tpsE do show a huge
difference in power consumption between systems with HDD vs SSD storage but
there are far too few toderive anything meaningful as far as CPU utilisation is
concerned:
http://www.tpc.org/tpce/results/tpce_energy _results.asp
obtained the TPC- E benchmark kit from Microsoft a fewyears ago& ran it in our lab
but found the workload characteristics tobe very CPU compute intensive & far less /
O intensive compared with TPC-C so the difference in power consumption between
the systems with / without SSDs is curious to me.
6 SeIectIng a Ircessr Ir SI Server 2012
Saturday 07 june 2014 09.22.42 ST http.//sqIperIrmance.cm/2013/01/system-cnIIguratIn/seIectIng-...
suspect that much of it comes down to CPU idle time rather than power differences
between the two types of storage, along the lines of your last comment - that the
CPUs get far more idle time when running on SSDs vs HDDs, yielding far less spinlock
time & context switching.
t's all very interesting & as electricity is getting more expensive, trading off a little CPU
performance for the long termpower savings ( which can cost bucket loads over a few
years) is becomming more attractive. Naturally you always need enough rawCPU
power toget the |ob done but trying to avoid power wastage isn't an easy assessment
in CPU selection.
Regards,
Creg Linwood
Reply
Cl enn Berry says:
anuary 10, 2013 at 12:36 AM
am all for using more efficient components where possible, both for
environmental reasons and for economic reasons. Despite that, think
mission critical database servers (which are usually relatively fewin
number at most organizations) are not the best place to make extreme
power efficiency a higher priority over CPU performance. The current
generation of servers from all of the ma|or vendors is much more efficient
than servers from a fewyears ago. Most 1U and 2U servers with Xeon E5
processors use between 200-400 watts, depending on how they are
configured and how much load they are under. n the United States,
electricity currently costs between 10- 15 cents per kWh in most areas.
f we gowith 400 watts, running 24 hours/day, that is 9.6 kWh per day. At
15 cents per kWh, that is $1.44 per day for the electricity. f you keep this
2U database server for three years, that would only be $1576.80 in
electricity for the lifetime of the server. That is relatively insignificant
compared tothe original hardware cost ($10, 000.00), the SQL Server 2012
core-licensing costs for 16-cores ($109,984.00), and the partial labor cost of
your staff over three years.
The extra CPU performance you get from choosing the correct processors
and disabling power management can be used for things like backup
compression and data compression, which may let you use a much smaller
number of spindles and storage space for your overall workload, which
would also translate intopower savings. Doing things like retiring older,
less efficient servers and consolidating or virtualizing onto fewer, new
database servers is a better strategy toreduce your electrical consumption.
t is a different story if you have hundreds or thousands of web servers
running open- source software. That would be a better example of where
power costs are a larger percentage of your overall costs.
Reply
Creg Li nwood says:
anuary 10, 2013 at 2:58 AM
Are the electricity costs you've quoted retail (home) rates or
data centre rates7 Data centre rates are usually higher than
retail rates as DC rates usually indirectly reflect aircon on top of
the actual per -server / rack draw and other factors such as
redundancy infrastructure.
n Australia, rack power is more like AUS$500 per kw/month in
a quality DC and a rack full of busy servers will typically be in
7 SeIectIng a Ircessr Ir SI Server 2012
Saturday 07 june 2014 09.22.42 ST http.//sqIperIrmance.cm/2013/01/system-cnIIguratIn/seIectIng-...
the vicinity of 10 kw, though less in racks with mixed servers
(SQL + other systems).
Power draw on our own servers vary from $100 > $500 per
month depending on the size of the server, how much internal
storage, howmany CPUs etc. The life power cost of our own
servers over 4 years runs into5 figure sums.
The power cost trend is also increasing & likely toincrease even
faster in future, particularly in countries like Australia that are
heavily taxing excessive power consumption. 'mnot sure how
power cost trends look in the US, the picture could be very
different of course.
'm alsonot suggesting that servers should be given any less
power than is necessary , |ust that power shouldn't be wasted
for both cost & environmental reasons & this is a factor in
choosing a CPU which might tip things more toward AMD. We
use both but haven't done any careful comparisons, hence
was interested in whether you might have done such a
comparison
Reply
6. Caurav Upadhy ay says:
anuary 15, 2013 at 6:45 AM
Thanks Clenn for putting together a great article!! have read your book ' SQL Server
hardware' and find it a great resource for everybody. ust wanted toknow when will
you be refreshing it with new and latest information.
Reply
7. Cl enn Thompson says:
anuary 29, 2013 at 8:50 PM
Creat article, now all have to dois to get the company tobuy the right servers and
not virtualize them. Thanks for the link.
Reply
8. Cary says:
uly 25, 2013 at 10:21 PM
see comments are a bit stale here, and in a causal search, this is surprisingly the only
spot turned up any real discussion on this topic. You would think it would merit a
little more for those in situations that aren't so clear-cut as towhere toallocate the db
dollars. With the price of SQL on the core licensing model , and the proliferation of CPU
configurations, there are a lot of trade-offs available. really like tooptimize things, so
maybe it's |ust me, but any SMB (like us) will benefit fromthe right blend.
We've been running a 200CB transactional database on SQL 2005 Std for about 3-1/2
years. We now run a mirror pair of HP C6 servers, each with a single X5550 and
36CB. ' ve got 20x 15K spindles on one data file, and performance is really fine for us,
excepting the bad query or relation that creeps in fromtime to time. CPU load is
typically 20% up to 30% peaks.
8 SeIectIng a Ircessr Ir SI Server 2012
Saturday 07 june 2014 09.22.42 ST http.//sqIperIrmance.cm/2013/01/system-cnIIguratIn/seIectIng-...
We recently decided tobite the bullet and upgrade to(2012) Enterprise Ed. Our goal
is continuous availability, so we expect to help meet that with the online index
rebuilds, table partitioning, and some of the Always-On features - and more memory.
(Right now index rebuilds are pretty much ignored, making us more susceptible to
blocking froma dirty query. And, while we can probably do the table partitioning by
other means, we expect it to be worthwhile in terms of app maintenance and |ust
plain time touse the Enterprise feature.)
While we're at it, we decided it would be prudent torefresh the hardware. We're not
going > 4 cores - there's clearly no need, and the pricey software addition makes that
even easier . Besides |ust having newhardware, the prime benefit there is more
memory (another prime benefit of Enterprise). Not having really looked at hardware
in awhile, was initially disappointed that most all the CPUs of any clock rate are 6-core
+. Now believe that you can disable cores in the HP bios (but haven' t confirmed),
and then alsonoticed the E5-2643 chip that would seem perfect for us. We don't
usually do CTO, but that would seem tobe a great fit. 192CB on one CPU in an HP
Cen8 is pretty cheap (~ $2K). don' t want to cover up bad database development or
admin with hardware, but performance headroomat that price is a no-brainer.
even notice a 2-core CPU on the Cen8 ( E5-2637) that would really let me go nuts with
the RAM. So, ' meven wondering if there might be a good trade-off there between
nearly 400CB of RAM, and fewer HDD spindles.
Perhaps you could say we're not performance constrained, but are more people
constrained. f a few thousand extra at refresh keeps the database humming, and
saves us time, it seems well worth it. This might not be a great case study for your
post, but thought 'd share it. ' malways open to comments in return. Thanks!
Reply
9. Cl enn Berry says:
uly 26, 2013 at 3:55 PM
Hi Cary ,
The article is several months old, but am glad you found it. The ntel Xeon E5-2643 is
an excellent choice tominimize your SQL Server 2012 core-license costs. A dual-core,
ntel Xeon E5-2637 would actually be a terrible choice, since you are forced tobuy a
minimum of four core licenses per physical processor . You also won't be able to use
more than 192CB of RAMif you only have one physical processor in a two- socket
server like a DL380p Cen 8. Having 192CB should be sufficient compared tothe
36CB you have now.
Reply
Cary Fost er says:
uly 26, 2013 at 4:23 PM
Thanks for your reply, Clenn. was |ust reviewing the SQL licensing this
morning, and saw the same thing you mentioned - that the license
requires a minimumof 4 cores per physical CPU. nterestingly , though, it
looks like running SQL in a vM only requires a minimum of 4 core licenses,
so could deploy on a dual 2643 box (or any 2P server), and have access
to the larger RAMcapacity.
Maybe it's not typically done, but when 've got a smaller database of fairly
consistent size, and can provision 256CB of RAMfor about the price of a
dozen 15K drives, then that's a trade-off that would seem tomake
eminent sense. With 768CB headroomin the Cen8 (when 32CB DMMs
become affordable) , this would be the last machine we' d ever need ,- )
9 SeIectIng a Ircessr Ir SI Server 2012
Saturday 07 june 2014 09.22.42 ST http.//sqIperIrmance.cm/2013/01/system-cnIIguratIn/seIectIng-...
The only thing to disrupt that perfect picture is running SQL on a
hypervisor . 'm generally not a fan of complicating things that way on line-
of- business platforms. Maybe 'm|ust old school. Or |ust old.
Thanks again for your reply , Clenn, and for helping sort out the options in
some sensible way .
Reply
10. Ceorge says:
February 10, 2014 at 7:06 PM
Why not a AMD cpu7
Reply
Cl enn Berry says:
February 10, 2014 at 7:16 PM
Because you will pay significantly more in SQL Server 2012 licensing costs
and get significantly lower single-threaded performance with modern AMD
processors compared tomodern ntel processors. That does not seem like
a good idea, and it may be difficult to|ustify froma business perspective.
Reply
Ceorge says:
February 10, 2014 at 7:18 PM
Thank you, Sir.
Reply
Aaron Bert rand says:
February 10, 2014 at 7:18 PM
Clenn's got a whole paragraph on that, no7
Also see:
http://sportstoday.us/technology/sql-server-2012-demystifying-hardware
-processor -vendor-selection.aspx
Notably :
"AMD processors have higher physical core counts in their processors
compared tontel , and they provide lower performance per physical core.
This forces you topay for more SQL Server 2012 core licenses but get lower
single- threaded performance, which is not a very good combination."
And here:
http://www.myotherpcisacloud.com/post/2013/04/19/Processor-Shopping-
for- SQL-Server- 2012.aspx
Notably :
10 SeIectIng a Ircessr Ir SI Server 2012
Saturday 07 june 2014 09.22.42 ST http.//sqIperIrmance.cm/2013/01/system-cnIIguratIn/seIectIng-...
Logo and site design SQL
Sentry, nc.
User contributions are
licensed under cc-sa with
attribution required.

Name
E- mail
UR
Should use NOT N, OUTER APPLY, LEFT OUTER ON, EXCEPT, or NOT
EXSTS7
Potential enhancements to
ASPState
"AMD cores tend toperform much worse per core than ntel cores on an
OLTP workload, and that crazy numbers of processor cores present with
diminishing returns regardless of the manufacturer . So far things are not
looking good for AMD. AMD can pack more cores on a die, but that |ust
simply does not make up for their gap in single-threaded performance."
Those are |ust a couple of quotes pulled out, please read the whole
articles ( and search for many others most of which come tosimilar
conclusions) .
Reply
Leav e a Repl y
( required)
( required)
Y our Comment
Submit Comment
Notify me of followup comments via e-mail. You can also subscribe without commenting.

You might also like