Relatively high CPU usage on measly 70 MB/s replication write throughput

[joonas-fi]

Reads amount to ~ 70 MBs, so do writes (symmetric because we write what we read). We are only doing CRC32 (IEEE) during replication:

varasto/pkg/stoserver/stodiskaccess/diskaccess.go

Lines 93 to 110 in 278ac87

	expectedCrc32, err := d.metadataStore.QueryBlobCrc32(ref)
	if err != nil {
	return fmt.Errorf("Replicate() QueryBlobCrc32: %v", err)
	}
	rawContent, err := fromDriver.RawFetch(ctx, ref)
	if err != nil {
	return err
	}
	defer rawContent.Close()
	crc32VerifiedReader := hashverifyreader.New(rawContent, crc32.NewIEEE(), expectedCrc32)
	if err := toDriver.RawStore(ctx, ref, crc32VerifiedReader); err != nil {
	return err
	}
	return d.metadataStore.WriteBlobReplicated(ref, toVolumeID)

CPU usage averages to 55 % across all cores. We should be expected to get 1,000 to 5,000+ MB/s per core on an AMD Ryzen 5 2400G (claimed by ChatGPT), so CPU usage really shouldn't be much above base load during replication.

There's something funky going on, my hunch is that it's the DB doing a write transaction per every blob replicated. 70 MB/s should equal to 18 blobs a second, therefore 18 write txns a second - assuming 4 MB blobs on average for my data set. This should be peanuts for any DB.

[joonas-fi]

If my hunch is correct, this should be mitigated by #206