The New Hotness

The discussion about scrubbing hard drives suggested I really should be using larger block sizes to wring better performance from the hardware.

So I ran variations on this theme:

time sudo dd if=/dev/urandom of=/dev/sdc bs=4K count=32K

For the BS (“block size”) parameter, 1K = 1024 and 1KB = 1000. Similarly for 1M vs. 1MB.

The results, viewed as a picture because WordPress seems unable to import a formatted spreadsheet from LibreOffice like it used to:

Each operation transfers 128 MB (128 x 2²⁰ = 131 x 10⁶) bytes. The variations probably come from other stuff going on, most notably the USB-to-serial adapter driving the plotter while I’m testing a tweak to the Superformula demo code.

Reads ever so much faster than writes, so the USB adapter definitely isn’t getting in the way; I assume the drive accepts the commands & data as fast as its little heads can carry them away. The data, being relentlessly pseudo-random, won’t get compressed along the way.

So, in round numbers, the block size just absolutely does not make any difference.

Update: Based on an early comment from Edward Berner to a previous post, I was looking in the wrong place:

dd if=/dev/urandom of=/dev/zero bs=4K count=32K
32768+0 records in
32768+0 records out
134217728 bytes (134 MB) copied, 9.63064 s, 13.9 MB/s
dd if=/dev/urandom of=test.bin bs=4K count=32K
32768+0 records in
32768+0 records out
134217728 bytes (134 MB) copied, 10.018 s, 13.4 MB/s
dd if=test.bin of=/dev/zero bs=4K count=32K
32768+0 records in
32768+0 records out
134217728 bytes (134 MB) copied, 0.0385358 s, 3.5 GB/s
dd if=test.bin of=test2.bin bs=4K count=32K
32768+0 records in
32768+0 records out
134217728 bytes (134 MB) copied, 0.45044 s, 298 MB/s

I installed an SSD on this box a while ago, so the 3.5 GB/s disk-to-discard speed represents the SSD’s read rate. The 298 MB/s disk-to-disk speed would be its write speed, probably with some clever buffering going on.

So the real bandwidth limitation in wiping a disk comes from the pseudo-random generator behind /dev/urandom, not the disk or USB interface. It would probably be faster to fill a 1 GB (or more) file with noise at 14 MB/s, then copy it enough times to fill the drive at whatever speed the drive can handle it.

Thanks, Edward, for figuring that out!