eMMC performance, reads slower than writes

[Loki]

I’m curious as to what people are seeing in terms of I/O performance from the internal eMMC, I’m seeing that reads are significantly slower than writes which is unexpected.

Writes:

purism@pureos:~$ dd if=/dev/zero of=io-test-file bs=4096 count=250000 conv=fdatasync
250000+0 records in
250000+0 records out
1024000000 bytes (1.0 GB, 977 MiB) copied, 16.0372 s, 63.9 MB/s

Reads:

purism@pureos:~$ dd if=io-test-file of=/dev/null bs=4096 count=250000 iflag=direct
250000+0 records in
250000+0 records out
1024000000 bytes (1.0 GB, 977 MiB) copied, 59.1235 s, 17.3 MB/s

These results are consistently repeatable.

The phone in question is running Byzantium luks variant, although bench marking the crypto would suggest that it’s capable of encryption/decryption magnitudes faster than disk I/O so shouldn’t be a factor?

[dos]

Shouldn’t you use oflag=direct in the first invocation to make both readings actually comparable?

[Loki]

Thanks, looks like that would be the case, I had thought that fdatasync was enough for it to be a direct write.

Running again with oflag=direct (would the conv=fdatasync still be required?) shows more or less equal read/write speeds…

purism@pureos:~$ dd if=/dev/zero of=io-test-file bs=4096 count=250000 conv=fdatasync oflag=direct
250000+0 records in
250000+0 records out
1024000000 bytes (1.0 GB, 977 MiB) copied, 51.3745 s, 19.9 MB/s

[Quarnero]

@Loki

I’m curious as to what people are seeing in terms of I/O performance from the internal eMMC

I prefer this result:

@spaetz

Here you go:

  WRITE: bw=62.1MiB/s (65.1MB/s), 62.1MiB/s-62.1MiB/s (65.1MB/s-65.1MB/s), io=248MiB (260MB), run=3993-3993msec
  READ: bw=110MiB/s (116MB/s), 110MiB/s-110MiB/s (116MB/s-116MB/s), io=248MiB (260MB), run=2249-2249msec

[Quarnero]

@Loki, I managed to find my post (I’m tired now to explain everything on how I’ve tested) yet creating test_eMMC_zero_data file of preferable size is important. Please follow this procedure that I made when I had more time, hope that helps:

@Quarnero

Today I’ve used count=2048 (perhaps 4096 would be ideal), yet quicker count of 512 will do as well. Note that for test_data files you’ll need su rights to remove those (sudo rm test_XX_data).

Afterwards, when ready to use above created file (within that folder) you might like to take a look (proceed) with something as posted inside of my post #40, same thread. All actually based on @amosbatto request, his post #24.

I’d also like to see your Librem 5 eMMC performance results (therefore this post), sure, if you don’t mind.

[irvinewade]

[MyL5] Australia/New Zealand

Redoing the test of the eMMC drive but using only dd and using a 600 MiB file … read: about 57 MB/s write: about 22 MB/s That’s better!

[Loki]

Thanks for the links. I’m trying not to get too caught up in the details these days, this caught my eye as I was taking a baseline to compare with after changing the encryption on the root file system from LUKS1 to LUKS2 and upping the key size, but as dos has pointed out, I needed the oflag=direct in order for the read and write results to be comparable.