On Mon, May 18, 2020 at 12:58:28PM +1000, Dave Chinner wrote: > > Topic: Version 3 log format > > Scope: Performance > Removing sector size limits > Large stripe unit log write alignment > > Proposal: > > The current v2 log format is an extension of the v1 format which was > limited to 32kB in size. The size limitation was due to the way that > the log format requires every basic block to be stamped with the LSN > associated with the iclog that is being written. > > This requirement stems from the fact that log recovery needed this > LSN stamp to determine where the head and tail of the log lies, and > whether the iclog was written completely. The implementation > requires storing the data written to the first 32 bits of each > sector of iclog data into a special array in the log header, and > replacing the data with the cycle number of the current iclog write. > When the log is replayed, before the iclog is read the data is > extracted from the iclog headers anre written back over the cycle > numbers so the transaction information is returned to it's original > state before decoding occurs. > > For V2 logs, a set of extension headers were created, allowing > another 7 basic blocks full of encoded data, which allows us to remap an > extra 7 32kB segments of iclog data into the iclog header. This is > where the 256kB iclog size limit comes from - it's 8 * 32kB > segments. > > As the iclogs get larger, this whole encoding scheme because more > CPU expensive, and it largely limits what we can do with expanding > iclogs. It also doesn't take into account how things have changed > since v2 logs were first designed. > > That is, we didn't have delayed logging. That meant iclogbuf IO was > the limiting factor to commit rates, not CPU overhead. We now do > commits that total up to 32MB of data, and we do that by cycling > through it iclogbuf at a time. As a result, CIL pushes are largely > IO bound waiting for iclogbufs to complete IO. Larger iclogbufs here > would make a substantial difference to performance when the CIL > is full, resulting in less blocking and fewer cache flushes when > writing iclogbufs. > > The question is this: do we still need this cycle stamping in every > single sector? If we don't need it, then a new format is much > simpler than if we need basic block stamping. > > From the perspective of determining if a iclog write was complete, > we don't trust the cycle number entirely in log recovery anymore. > Once we have the log head and the log tail, we do a CRC validation > walk of the log to validate it. Hence we don't really need cycle > data in the log data to validate writes were complete - the CRC will > fail if a iclogbuf write is torn. > > So that comes back to finding the head and tail of the log. This is > done by doing a binary search of the log based reading basic blocks > and checking the cycle number in the basic block that was read. We > really don't need to do this search via single sector IO; what we > really want to find is the iclog header at the head and the tail of > the log. > > To do this, we could do a binary search based on the maximum > supported iclogbuf size and scan the buffers that are read for > iclog header magic numbers. There may be more than one in a buffer, > (e.g. head and tail in the same region) but that is an in-memory > search rather than individual single sector IO. Once we've found an > iclog header, We can read the LSN out of the header, and that tells > us the cycle number of that commit. Hence we can do the binary > search to find the head and tail of the log without needing have the > cycle number stamped into every sector. > > IOWs, I don't see a reason we need to maintain the per-basic-block > cycle stamp in the log format. Hence by removing it from the format > we get rid of the need for the encoding tables, and we remove the > limitation on log write size that we currently have. Essentially we > move entirely to a "validation by CRC" model for detecting > torn/incomplete log writes, and that greatly reduces the complexity > of log writing code. > > It also allows us to use arbitrarily large log writes instead of > fixed sizes, opening up further avenues for optimisation of both > journal IO patterns and how we format items into the bios for > dispatch. We already have log vector buffers that we hand off to the > CIL checkpoint for async processing; it is not a huge stretch to > consider mapping them directly into bios and using bio chaining to > submit them rather than copying them into iclogbufs for submission > (i.e. single copy logging rather than the double copy we do now). > And for DAX hardware, we can directly map the journal.... > > But before we get to that, we really need a new log format that > allows us to get away from the limitations of the existing "fixed > size with encoding" log format. > > Discussion: > - does it work? > - implications of a major incompat log format change > - implications of larger "inflight" window in the journal > to match the "inflight" window the CIL has. Giant flood of log items overwhelming the floppy disk(s) underlying the fs? :P > - other problems? > - other potential optimisations a format change allows? Will have to ponder this in the morning. > - what else might we add to a log format change to solve > other recovery issues? Make sure log recovery can be done on any platform? --D > > -- > Dave Chinner > david@xxxxxxxxxxxxx
