Hi Nikolaus,

[...] is there a reason to not just bump the hardcoded buffer size?

Adding an option typically means that 50% of the people go with the
default (even when it's sub-optimal), 40% pick a value that's even
worse, and 10% actually benefit. So I'd like to avoid options whenever
possible.[...]
If changing the hard-coded BUFSIZE is an option, that's what I would prefer, too. That would increase the required memory for S3QL, tho. So when low-memory systems are a of any concern, increasing the BUFSIZE would be bad for them.

In this use case I use S3QL as a target for Bareos backups. Each backup is one single file that can get hundreds of GB big. Thus I chose a max-obj-size of 3GB, no compression (Bareos does that already) and a cache size of 100 GB. The file systems look like these (cache gets dropped between backup windows):

Directory entries:    4735
Inodes:               4737
Data blocks:          6211
Total data size:      6.35 TB
After de-duplication: 6.35 TB (100.00% of total)
After compression:    6.35 TB (99.94% of total, 99.94% of de-duplicated)
Database size:        1.91 MiB (uncompressed)
Cache size:           0 bytes, 0 entries
Cache size (dirty):   0 bytes, 0 entries
Queued object removals: 0

So there are relatively few objects/data blocks but they are 1GB on average. This is quite a different use case as than the default max-obj-size of 10MB. Before bumping the BUFSIZE we definitely should benchmark with the default max-obj-size, too.

Looking at contrib/benchmark.py, I can probably change this to also benchmark different BUFSIZEs for the upload so that we can get some data from different configurations.


References:
https://github.com/python/cpython/commit/4f1903061877776973c1bbfadd3d3f146920856e increased buffer from 16KB to 64KB https://blogs.blumetech.com/blumetechs-tech-blog/2011/05/faster-python-file-copy.html 10 MB buffer for large files

--
You received this message because you are subscribed to the Google Groups 
"s3ql" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
To view this discussion on the web visit 
https://groups.google.com/d/msgid/s3ql/3eb1cf7f-de64-639a-09b6-23333d06ec63%40jagszent.de.
For more options, visit https://groups.google.com/d/optout.

Reply via email to