That's 18,369MB per second from a Windows box... stuffed with $100 CPUs from 2009... and $75 HBAs.
Last edited:
How's this for Windows disk IO
- Thread starter dba
- Start date
Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.
That's 18,369MB per second from a Windows box... stuffed with $100 CPUs from 2009... and $75 HBAs.
This was from a Windows setup on my "Dirt Cheap Data Warehouse" (DCDW) server hardware - the HP DL585 G7 in my signature. I was running some pretty massive database queries but getting slower than expected results. It turns out that a Kettle ETL process that I was also running was borked and errantly running 32 of its own disk threads. The combined IO was a "personal best" for real world IO on this machine, and I doubt that I'll be able to top it without going MP Xeon E5.
Links:
http://forums.servethehome.com/diy-...warehouse-database-storage-servers-cheap.html
The Dirt Cheap Data Warehouse Architecture
Counting Triangles Even Faster - and Far Cheaper
The first two links above talk about the DCDW. The last one compares this DIY monster it to some $500K commercial database machines. The DCDW fares pretty well because... cheap SSD drives.
Links:
http://forums.servethehome.com/diy-...warehouse-database-storage-servers-cheap.html
The Dirt Cheap Data Warehouse Architecture
Counting Triangles Even Faster - and Far Cheaper
The first two links above talk about the DCDW. The last one compares this DIY monster it to some $500K commercial database machines. The DCDW fares pretty well because... cheap SSD drives.
Last edited:
Maybe his cpus have more memory channels than yours? $100 cpus from 2009 would likely mean socket 1366 triple channel, if you tested on a dual channel it would make sense.Awesome and how? After running ram disks to test rdma, I was only hitting 6-9GB per second, that is very impressive! Might forgo setting up FhGFS over my 40gb infiniband network.
You can RDMA gpus apparently, though its locked to the tesla models at the moment. It would be awesome to use a videocard from another machine to play games, I don't think the 0.1ms latency would be a problem if set up properly.
As for remote desktop, DP 1.2 is only ~18Gb and supports 4k@60hz, so plenty of headroom there.
If you want a fun experiment, try this: Run the STREAM benchmark with different memory configurations - one DIMM, two, three, etc. You'll be surprised at how slow a good sever will run when the RAM slots are nearly empty. Also, if you need speed from your RAM disks, try setting up multiple smaller disks instead of one big one - but you probably already read that in the forums.Awesome and how? After running ram disks to test rdma, I was only hitting 6-9GB per second, that is very impressive! Might forgo setting up FhGFS over my 40gb infiniband network
These are AMD 6174 CPUs. Four CPUs times four memory channels each equals >100GB/s memory bandwidth!
Last edited:
what if you change a quadro 6000 into a tesla? That should be easy to strap. I noticed it said if you have IOMMU you can't do GPUDIRECT.
GPUDIRECT is Tesla -> Mellanox-CX3 (only connect-x3) but the main gain of going from GPU -> NIC without passing through CPU is not going to work out so hot if its inside hypervisor! so it has to be ring-0 (not going to happen) or bare metal.
I am thinking -> what if you had USB/thunderbolt video card over network?
GPUDIRECT is Tesla -> Mellanox-CX3 (only connect-x3) but the main gain of going from GPU -> NIC without passing through CPU is not going to work out so hot if its inside hypervisor! so it has to be ring-0 (not going to happen) or bare metal.
I am thinking -> what if you had USB/thunderbolt video card over network?