93e28d661e
= Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent |
||
---|---|---|
.. | ||
callbacks | ||
cmd | ||
include | ||
tests | ||
Makefile.am |