Am 26.05.2017 um 10:57 hat Denis V. Lunev geschrieben: > On 05/26/2017 11:11 AM, Kevin Wolf wrote: > > It seems much clearer to me that simply unifying the three write > > requests into a single one is an improvement. And it's easy to do, I > > even had a patch once to do this. The reason that I didn't send it was > > that it seemed to conflict with the data cache approach > These changes help a lot on 2 patterns: > - writing 4Kb into the middle of the block. The bigger the block size, > more we gain > - sequential write, which is not sector aligned and comes in 2 requests: > we perform allocation, which should be significantly faster than actual > write and after that both parts of the write can be executed in parallel. > At my opinion both cases are frequent and important.
Both cases are helped somewhat by reducing the number of I/O requests. So as a first step, I really think we need to get Berto's patches ready to be merged. As for this series, there is one important restriction that you don't mention: Both cases only benefit from the changes when there is no backing file (or the backing file is sparse at the given cluster). The main reason why people use qcow2 is for snapshots, so qcow2 without or with an empty backing file is actually not my primary concern. To be honest, I am concerned a bit by the additional complexity brought in by this series (which also seemed to be more tacked on rather than fully integrated at the first sight), but I'll have to study it in more detail. Maybe another thing to keep in mind is that the sequential write case can completely avoid COW in theory. I wonder whether we should make another attempt at this. Previously, it was the complexity of my own prototypes that stopped me from actually merging them. Kevin
