On Monday 09 April 2012, Minchan Kim wrote: > 2012-04-07 오전 1:16, Arnd Bergmann 쓴 글: > > > larger chunks would generally be helpful, in order to guarantee that we > > the drive doesn't do any garbage collection, we would have to do all writes > > > And we should guarantee for avoiding unnecessary swapout, even OOM killing. > > > in aligned chunks. It would probably be enough to do this in 8kb or > > 16kb units for most devices over the next few years, but implementing it > > for 64kb should be the same amount of work and will get us a little bit > > further. > > > I understand it's best for writing 64K in your statement. > What the 8K, 16K? Could you elaborate relation between 8K, 16K and 64K? >From my measurements, there are three sizes that are relevant here: 1. The underlying page size of the flash: This used to be less than 4kb, which is fine when paging out 4kb mmu pages, as long as the partition is aligned. Today, most devices use 8kb pages and the number is increasing over time, meaning we will see more 16kb page devices in the future and presumably larger sizes after that. Writes that are not naturally aligned multiples of the page size tend to be a significant problem for the controller to deal with: in order to guarantee that a 4kb write makes it into permanent storage, the device has to write 8kb and the next 4kb write has to go into another 8kb page because each page can only be written once before the block is erased. At a later point, all the partial pages get rewritten into a new erase block, a process that can take hundreds of miliseconds and that we absolutely want to prevent from happening, as it can block all other I/O to the device. Writing all (flash) pages in an erase block sequentially usually avoids this, as long as you don't write to many different erase blocks at the same time. Note that the page size depends on how the controller combines different planes and channels. 2. The super-page size of the flash: When you have multiple channels between the controller and the individual flash chips, you can write multiple pages simultaneously, which means that e.g. sending 32kb of data to the device takes roughly the same amount of time as writing a single 8kb page. Writing less than the super-page size when there is more data waiting to get written out is a waste of time, although the effects are much less drastic as writing data that is not aligned to pages because it does not require garbage collection. 3. optimum write size: While writing larger amounts of data in a single request is usually faster than writing less, almost all devices I've seen have a sharp cut-off where increasing the size of the write does not actually help any more because of a bottleneck somewhere in the stack. Writing more than 64kb almost never improves performance and sometimes reduces performance. >From the I've done, a typical profile could look like Size Throughput 1KB 200KB/s 2KB 450KB/s 4KB 1MB/s 8KB 4MB/s <== page size 16KB 8MB/s 32KB 16MB/s <== superpage size 64KB 18MB/s <== optimum size 128KB 17MB/s ... 8MB 18MB/s <== erase block size > > I'm not sure what we would do when there are less than 64kb available > > for pageout on the inactive list. The two choices I can think of are > > either not writing anything, or wasting the swap slots and filling > > > No wrtite will cause unnecessary many pages to swap out by next prioirty > of scanning and we can't gaurantee how long we wait to queue up to 64KB > in anon pages. It might take longer than GC time so we need some deadline. > > > > up the data with zeroes. > > > Zero padding would be a good solution but I have a concern on WAP so we > need smart policy. > > To be honest, I think swapout is normally asynchonous operation so that > it should not affect system latency rather than swap read which is > synchronous operation. So if system is low memory pressure, we can queue > swap out pages up to 64KB and then batch write-out in empty cluster. If > we don't have any empty cluster in low memory pressure, we should write > out it in partial cluster. Maybe it doesn't affect system latency > severely in low memory pressure. The main thing that can affect system latency is garbage collection that blocks any other reads or writes for an extended amount of time. If we can avoid that, we've got the 95% solution. Note that eMMC-4.5 provides a high-priority interrupt mechamism that lets us interrupt the a write that has hit the garbage collection path, so we can send a more important read request to the device. This will not work on other devices though and the patches for this are still under discussion. > If system memory pressure is high(and It shoud be not frequent), > swap-out B/W would be more important. So we can reserve some clusters > for it and I think we can use page padding you mentioned in this case > for reducing latency if we can queue it up to 64KB within threshold time. > > Swap-read is also important. We have to investigate fragmentation of > swap slots because we disable swap readahead in non-rotation device. It > can make lots of hole in swap cluster and it makes to find empty > cluster. So for it, it might be better than enable swap-read in > non-rotation devices, too. Yes, reading in up to 64kb or at least a superpage would also help here, although there is no problem reading in a single cpu page, it will still take no more time than reading in a superpage. > >>> 2) Make variable sized swap clusters. Right now, the swap space is > >>> organized in clusters of 256 pages (1MB), which is less than the typical > >>> erase block size of 4 or 8 MB. We should try to make the swap cluster > >>> aligned to erase blocks and have the size match to avoid garbage collection > >>> in the drive. The cluster size would typically be set by mkswap as a new > >>> option and interpreted at swapon time. > >>> > >> > >> If we can find such big contiguous swap slots easily, it would be good. > >> But I am not sure how often we can get such big slots. And maybe we have to > >> improve search method for getting such big empty cluster. > > > > As long as there are clusters available, we should try to find them. When > > free space is too fragmented to find any unused cluster, we can pick one > > that has very little data in it, so that we reduce the time it takes to > > GC that erase block in the drive. While we could theoretically do active > > garbage collection of swap data in the kernel, it won't get more efficient > > than the GC inside of the drive. If we do this, it unfortunately means that > > we can't just send a discard for the entire erase block. > > > Might need some compaction during idle time but WAP concern raises again. :( Sorry for my ignorance, but what does WAP stand for? Arnd -- To unsubscribe from this list: send the line "unsubscribe linux-mmc" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
