On Wed, Feb 26, 2020 at 9:20 AM Jan Kara <jack@xxxxxxx> wrote: > > On Wed 26-02-20 08:46:42, Dan Williams wrote: > > On Wed, Feb 26, 2020 at 1:29 AM Jonathan Halliday > > <jonathan.halliday@xxxxxxxxxx> wrote: > > > > > > > > > Hi All > > > > > > I'm a middleware developer, focused on how Java (JVM) workloads can > > > benefit from app-direct mode pmem. Initially the target is apps that > > > need a fast binary log for fault tolerance: the classic database WAL use > > > case; transaction coordination systems; enterprise message bus > > > persistence and suchlike. Critically, there are cases where we use log > > > based storage, i.e. it's not the strict 'read rarely, only on recovery' > > > model that a classic db may have, but more of a 'append only, read many > > > times' event stream model. > > > > > > Think of the log oriented data storage as having logical segments (let's > > > implement them as files), of which the most recent is being appended to > > > (read_write) and the remaining N-1 older segments are full and sealed, > > > so effectively immutable (read_only) until discarded. The tail segment > > > needs to be in DAX mode for optimal write performance, as the size of > > > the append may be sub-block and we don't want the overhead of the kernel > > > call anyhow. So that's clearly a good fit for putting on a DAX fs mount > > > and using mmap with MAP_SYNC. > > > > > > However, we want fast read access into the segments, to retrieve stored > > > records. The small access index can be built in volatile RAM (assuming > > > we're willing to take the startup overhead of a full file scan at > > > recovery time) but the data itself is big and we don't want to move it > > > all off pmem. Which means the requirements are now different: we want > > > the O/S cache to pull hot data into fast volatile RAM for us, which DAX > > > explicitly won't do. Effectively a poor man's 'memory mode' pmem, rather > > > than app-direct mode, except here we're using the O/S rather than the > > > hardware memory controller to do the cache management for us. > > > > > > Currently this requires closing the full (read_write) file, then copying > > > it to a non-DAX device and reopening it (read_only) there. Clearly > > > that's expensive and rather tedious. Instead, I'd like to close the > > > MAP_SYNC mmap, then, leaving the file where it is, reopen it in a mode > > > that will instead go via the O/S cache in the traditional manner. Bonus > > > points if I can do it over non-overlapping ranges in a file without > > > closing the DAX mode mmap, since then the segments are entirely logical > > > instead of needing separate physical files. > > > > Hi John, > > > > IIRC we chatted about this at PIRL, right? > > > > At the time it sounded more like mixed mode dax, i.e. dax writes, but > > cached reads. To me that's an optimization to optionally use dax for > > direct-I/O writes, with its existing set of page-cache coherence > > warts, and not a capability to dynamically switch the dax-mode. > > mmap+MAP_SYNC seems the wrong interface for this. This writeup > > mentions bypassing kernel call overhead, but I don't see how a > > dax-write syscall is cheaper than an mmap syscall plus fault. If > > direct-I/O to a dax capable file bypasses the block layer, isn't that > > about the maximum of kernel overhead that can be cut out of this use > > case? Otherwise MAP_SYNC is a facility to achieve efficient sub-block > > update-in-place writes not append writes. > > Well, even for appends you'll pay the cost only once per page (or maybe even > once per huge page) when using MAP_SYNC. With a syscall you'll pay once per > write. So although it would be good to check real numbers, the design isn't > non-sensical to me. True, Jonathan, how many writes per page are we talking about in this case?
