Hello together!

I've a system with 8 SSDs (each ~500MByte/s read/write) connected to a LSI 9271-8i RAID-controller as a RAID0 array.
Under Windows I get write and read rates of about 4GByte/s.

Under Linux I test for example with
time dd if=/dev/zero of=/dev/sdX bs=1M count=8000
but then I only get this:
8000+0 records in
8000+0 records out
8388608000 bytes (8.4 GB) copied, 7.61166 s, 1.1 GB/s

real 0m7.613s
user 0m0.000s
sys 0m7.572s

So the rate is too low and the reason seems to be the CPU utilization because dd is running on one core and its "system utilization" is at 100% during the whole runtime.
If I run 12 dd in parallel and write to the same device then the aggregated rates are ~3GByte/s.
The utilization of the benchmark bonnie++ is also ~100%.

Hence I've these questions:

• Is there a way to get rid of this CPU-limitation during reads/writes?
  I mean: The system has 2 processors with each 6 cores. So if dd uses write() or a similar call, is there no "paralelized version" under linux?
• Is there a benchmark which generates multiple threads/processes for concurrent reads/writes?

Many thanks for any help!

dd is a poor way to test it no matter how you set it up.

Use some other method.


Might look at how you set up the raid, lsi driver, kernel settings, swap file or swap partition too, filesystem may be most important too for the size of files you are trying to move. We'd need to know some of that.

Might look at iozone too.

I'd welcome others thoughts on this too.

What software are you using in Windows to do your testing? Try getting a Linux equivalent in order to get a more accurate reading.

@jefro
I write directly to the device and I don't use any filesystem.
And I know that I must use another method. That's why I've created this thread.
The kernel is one from debian but it doesn't matter. I get same results with ubuntu kernels.

@dman58
Under Windows I use "ATTO Disk Benchmark".
And my second question is about an "equivalent benchmark" because I don't know one for linux that isn't CPU-limited on such fast drives.

Use a multi-threaded capable tool - fio, dbench (maybe), iozone ...

Phoronix does truckloads of benchmarks, and tell you what they use - might be a good place to start looking.