Best CPU for database work?

[Skiny]

So, as the title says, I'm wondering what the best options are when it comes to processors best suited to working with databases?

I'm running software that processes around 8,000 (fairly complex) search queries per minute, 24/7, on an i7 930 CPU. I get fairly high CPU usage (to my surprise ), and as the number of queries is only going to go up over the coming months, I was thinking about maybe getting a new CPU(s).

Are there any specific CPUs that you guys would recommend to deal with this type of task? I use the computer for heavy multi-tasking, as well as the database work.

 

[Arkaign]

If you're not an OC'er, your answer may lie in the 4770 (non-K), due to very good IPC. Not many database applications are actually all that efficient at loading many cores, so you want per-core performance to be tops.

You want to pair that with heaps of ram, with prices as they are, getting 16GB or more of high-speed memory (at least good reviewed DDR3-1866!), along with a very fast disk system. The better Sata controllers on the Haswell mobos should help in this regard. If you can fit it all, moving all of your database data to a pair of 500GB 840 Pros would be nice. There are even higher-end professional SSDs out there, but they run into the thousands of dollars per drive in a hurry.

The answer really is : how much money do you have? Hahah

 

[Vectronic]

(Disclaimer: no real experience) But I'd agree with that -^

But, really it sounds like you should get another computer, leave the 930 as the DB, get a 3570/4670 for your other needs. Stripping out all the non-DB stuff, less interruptions, etc would probably cope fine if it's working currently as part desktop part server.

 

[TerryMathews]

So, as the title says, I'm wondering what the best options are when it comes to processors best suited to working with databases?

I'm running software that processes around 8,000 (fairly complex) search queries per minute, 24/7, on an i7 930 CPU. I get fairly high CPU usage (to my surprise ), and as the number of queries is only going to go up over the coming months, I was thinking about maybe getting a new CPU(s).

Are there any specific CPUs that you guys would recommend to deal with this type of task? I use the computer for heavy multi-tasking, as well as the database work.

What database program?

 

[Cerb]

Just going to an i7-4770 should offer a 30% or greater improvement, for CPU-limited queries. With a DBMS that scales by thread well, like MySQL, you could be looking at up to 100%, depending on query and storage engine. But, if not 30-50%, then surely there's some other areas needing optimization.

The improvements in cores over the last few generations have been really good to MySQL and MS SQL (includes Access, now, but that won't load up the cores), and probably others, even more-so than to more commonly-looked-at things like games.

How much of an improvement it might be will be highly dependent on currently-unknown factors, including the DBMS, DB config, storage config (if a bottleneck), and queries themselves. However, the i7-4770 and Xeon E3-1240V3 (especially if you use a video card), are solid, and don't involve high-end spending (<=$100 more than if building any other midrange PC).

 

[ph2000]

first thing first : budget

what kind of query are you making ?
is it optimized already ? memory usage ?

 

[BrightCandle]

Very much depends on the software. But most major database programs use a lot of threads pretty well, its not uncommon to get a 2 socket Xeon board and fill it with 12 core a socket and see definite benefits from doing so. But it can really depend on the queries, how complex they are, how much is happening in parallel, how many clients etc. Some queries work much better with clock speed/single core performance and some really benefit from parallel cores. It normally requires good performance measurements and statistics to get a true picture of what hardware is likely to bring the most performance.

Thus in this case I would probably avoid the standard consumer 4 core CPUs and start looking for more cores. X79 and a 6 core if that is sufficient, otherwise consider getting into the server hardware world and either go single socket with many cores or really start to get into the hardware game with dual socket. But I also wouldn't do a thing without looking at the usage of the machine today, profiling the queries and getting a picture of how its all being optimised currently. Its easier for you than most, you already have an application in front of you that you can build hardware to, most people buying DB hardware are stuck guessing how they will use it!

 

[cytg111]

Of that high CPU usage, how much is kernel time?
Off the bat I'd say a CPU with a large amount of cache, but that is still a pretty uneducated guess! What kind of data(how big) is these queries running against, and how complicated are they?

in essense, you need to define your bottleneck, is it CPU (and L1, L2 cache bound), RAM, disk io etc.

 

[caution]

There are special productivity/DB benchmarks out there. I think AMD cpus do a great job at 'em

 

[Skiny]

Thanks very much to everyone for all of the replies.

CPU usage seems fairly consistent across all 4 cores of my current CPU, so I would guess extra ones would be beneficial. Usage seems to range from 20% to 90+%. A few people have mentioned the 4770.

The entire database is currently stored in RAM, so I don't believe their are any other bottlenecks. RAM usage is also fine <50% mostly. It seems it's entirely a CPU related issue.

I've thought of using a separate computer for the work, although things would be much more convenient if everything could be contained within one. Plus I (hope), costs would be the same or less than buying/running two.

The database software is RavenDB and it does seem to support multiple CPUs, from what I've read. I'm not sure how much of an improvement would be gained with two CPUs though, over say a 6-core i7. It's hard to judge.

The queries are with mostly with many keywords - maybe 40 per search - with added search commands. Also, multiple other fields are also selected for each search. I'm not really familiar enough to say if that's considered a 'complex query' or not.

Not sure what to make of AMD these days. More cores for sure, but I've not heard anything good about AMD for some time.

 

[Cerb]

I can't find too much on AMD, and I wouldn't trust AMD to be better w/o some data. Intel is certainly easier to speculate at. I can't find anything one way or another about Hyperthreading, but it shouldn't hurt, and your current CPU has it (so if it's helping you out now, it would be silly to get a new CPU w/o it).

With a 6-core i7, you get much more L3, but without spending for the 3970, the benefits over a LGA1150 i7 are dubious, with the 3930 being 2 CPU generations behind, and your not being RAM-limited. On the i7-4770K is often dead even with data-crunching tasks, and not far behind when not.

http://www.anandtech.com/bench/Product/552?vs=836
Note how little it's faster, when it is.

Things like rendering are often bandwidth-limited, so get quite a bit more out of LGA2011, with its 2 more RAM channels and 7MB more L3 cache. For your application, that appears to be a big question mark, since a Core i3-2100 is the fastest CPU I've found tested at all, and it was for specific person's usage involving supposed performance improving patches, last year. An LGA2011 Core i7 might be significantly faster, might be slightly faster, or might not be any faster, than a LGA1150 Core i7. But, either one should be a good bit faster than what you're running.

 

[Jodell88]

If you can figure out how to use the website, Openbenchmarking.org should have what you need.

 

[crashtech]

If the app is maxing out 8 threads, chances are it could use more. How would one determine this based on type of usage? Might be good to ask RavenDB:

https://groups.google.com/forum/#!forum/ravendb

IB-E is just around the corner.

 

[TerryMathews]

I can't get the post to quote right.

You said the whole DB is in system RAM. Have you considered the possibility that you're bandwidth limited? Could a multi - socket system with NUMA help?