RAMFS seems slower than expected — what is the bottleneck?

Question

RAMFS doesn't uses swap (TMPFS does). So I mount the FS

cd /mnt/; mkdir SOMETHING; mount -t ramfs -o size=1500m ramfs /mnt/SOMETHING; cd /mnt/SOMETHING

I generate a testfile:

dd if=/dev/zero of=testfile_500MB bs=524288000 count=1

I cp the test file (from memory to memory)

time cp testfile_500MB testfile_500MB_cptestfile

real    0m0.599s
user    0m0.008s
sys     0m0.592s

I have one DDR2-1066 module in my PC. What is the bottleneck? According to this wikipedia article:

http://en.wikipedia.org/wiki/List_of_device_bandwidths

it would need to do more then 10 GByte/sec! It's waay not more 10 GByte/sec

500 MByte in 0.599 sec = 835 MByte/sec. It's slow.

Is the CPU the bottleneck? OR what? :O (E7300 @ 2.66GHz)

UPDATE: is measuring a "cp" with "time" a good thing?

Answer 1

Your test probably isn't long enough to average out the overhead of running cp, so I don't know if that's a good test. You might want to try something like bonnie++.

Still, the number you came up with doesn't seem unreasonable to me. If memtest86+ is to be believed, most systems with dual-channel RAM will do 2-3GB/s to main memory. Single-channel (as you have with only one stick of RAM) is going to be less (but not necessarily half). Subtract some understandable overhead, and a bit less than 1GB/s sound plausible.

Answer 2

Using cp, the data will go through several buffers (user side and kernel side), so the memory bandwidth required can be multiplied by a big factor. Maybe ramfs internals will also add some overhead. So, cp is definitely not a good tool to test memory bandwidth.

There must be some specific tools lying around for that.

Answer 3

Actually it's even slower than you think. If you create a file in this way it should be sparse and by default filesystems are mounted with asynchronous I/O.

Certainly cp is well known for being rather slow at copying large files (there are various ways to address this - the most common being to use cpio across a pipe).

What is it you are trying to measure the speed of? If you want to measure your memory I/O, use memtest. If you want a realistic measure of ramfs performance, use fio