Fastest 4 way Database Server

  • I'm looking to see what I would replace our 4 way Xeon X7350 2.93 GHz ( 16 cores ) cluster with if I wanted an upgrade.

    I have some obvious points:-

    1.more bandwidth across cpu and memory

    2.faster memory

    3.faster cpu clock speed

    Our current servers appear to struggle once they hit 2 million pages reads/sec ( or so ) so the ability to support more in memory reads quicker would be good.

    More cores isn't probably the answer as in general it's only the backups and maint jobs which make use of multiple cores. I know for instance that a 40 core server will produce a backup in 25% of the time as 8 cores, incidentally using 8 of the 40 cores to do the backup actually takes much longer than my current server.

    Initial research seems to indicate a max clock speed of 2.4GHz for a typical DL580 or R9000 , anyone know of faster servers or a set of respected benchmarks which compare servers ( and not by parallel processing which is unlikely to give me more performance )

    thanks.

    [font="Comic Sans MS"]The GrumpyOldDBA[/font]
    www.grumpyolddba.co.uk
    http://sqlblogcasts.com/blogs/grumpyolddba/

  • Really surprised that more cores reduce backup time as usually you are bound by I/O performance, even with an EVA with 100+ disks, if you have large databases.

    We use hyperbac (no plug intended) which works the cores harder than usual but we see significant improvements in backup performance and obviously nice and small backup files.

  • ah, sorry should have said we were using Litespeed which allows you to define the number of cores/threads for the backup.

    My backup using 8 threads on a 16 core 2.93GHz zeon typically takes 7 mins. Using 8 cores on a 40 core 2.4GHZ zeon took about 18 mins, but 40 cores took barely 2 mins ( same database and same config of litespeed except cores )

    My server has 16GB bandwidth to a dedicated san we use as DAS. The other server was using raided fusion io cards. In fact we can improve backup times further by writing to multiple files - our storage systems are reasonable.

    [font="Comic Sans MS"]The GrumpyOldDBA[/font]
    www.grumpyolddba.co.uk
    http://sqlblogcasts.com/blogs/grumpyolddba/

  • seems the top contender is an IBM 3850 X5.

    [font="Comic Sans MS"]The GrumpyOldDBA[/font]
    www.grumpyolddba.co.uk
    http://sqlblogcasts.com/blogs/grumpyolddba/

  • We use some DL580's and have a couple of Superdomes which are discontinued (i think).

    Try SGI who do the worlds fasted Windows Server.

    http://www.sgi.com/products/servers/altix/uv/[/url]

    These are in the running for a potential Superdome replacement.

  • thanks - I don't really need cores, 16 is ample, what I want is faster - so memory, bus speed and cpu clock speed etc.

    These look interesting, I currently use DL580s and Dell R9000, the burning question is if I were to upgrade what box do I choose? Other than taking the memory to 128GB or 192GB I'm not 100% sure - so thanks for another contender to add to the list.

    [font="Comic Sans MS"]The GrumpyOldDBA[/font]
    www.grumpyolddba.co.uk
    http://sqlblogcasts.com/blogs/grumpyolddba/

  • If you're only having issues after 16GB/s of throughput, you're in rarefied air already. You'll need to focus on RAM architecture, memory timings (critical), and bandwidth settings more than the machine itself.

    If you're really RAM bound, have you tried data_compression settings to see if you perhaps you have enough CPU to spare to up your overall throughput?

    I'd see if you can ask HP, Dell, and IBM to provide their best offers after explaining your situation, and try them with your workload.

    Another interesting thought: if you're not starved for CPU, consider benchmarking a single modern 10 or 12 core CPU; it's theoretically possible that if all the RAM is being directly controlled by the one CPU you'll see some overall performance improvement compared to a multi-socket setup.

  • no I don't have issues with bandwidth to the SAN, the issue is really the server architecture which seems to stall at around 2 million page reads a sec ( e.g. in memory io )

    I'm not convinced that a 2.26GHz cpu single core operation will match or be quicker than an existing 2.93GHz cpu core; e.g. faster single threaed operations require a quicker cpu - more cores only help if the operations are parallel. The existing box ticks along at 20 - 40 % cpu on 64GB ram with a dedicated SAN. there's no bandwidth issues, we have 16GB as we were running a series of tests prior to go live and using 4 HBAs per node seemed to give the best overall performance even though we didn't get to anywhere near that limit.

    So much of the benchmarks/published test results rely on increased parallel operations not on raw performance - it's all very frustrating!

    [font="Comic Sans MS"]The GrumpyOldDBA[/font]
    www.grumpyolddba.co.uk
    http://sqlblogcasts.com/blogs/grumpyolddba/

  • I meant 16GB/s RAM bandwidth (2 million pages/sec * 8KB/page).

    Take a look at the Intel E7-4870 and E7-8837, both of which have a "turbo" frequency of 2.8Ghz when the workload is light enough to allow for it, and a significantly newer architecture than previous chips.

  • ask HP, Dell, and IBM to provide their best offers after explaining your situation, and try them with your workload

    This looks like good advice to me. But I would be surprised if you get much more than a 10% to 20% difference at most. If a vendor could get 50% or 100% more memory throughput than the rest we would probably all know about it and be buying their kit. Someone may be willing to pick the best chips out of the production line (at a cost) to give you some extra boost, but if a 20% improvement compared with your current throughput will not keep the application happy for the next year or two then this looks like a dead end.

    You may need to look at other solutions that do not rely on a significant improvement in hardware throughput on a single box. That means splitting the workload over multiple machines, and you probably know better than me how to make that work with high workloads.

    Original author: https://github.com/SQL-FineBuild/Common/wiki/ 1-click install and best practice configuration of SQL Server 2019, 2017 2016, 2014, 2012, 2008 R2, 2008 and 2005.

    When I give food to the poor they call me a saint. When I ask why they are poor they call me a communist - Archbishop Hélder Câmara

Viewing 10 posts - 1 through 9 (of 9 total)

You must be logged in to reply to this topic. Login to reply