I think it would be beneficial for some Ignite users if we added such a partition warmup method to the public API. The method should be well-documented and state that it may invalidate existing page cache. It will be a very effective instrument until we add the proper scan ability that Vladimir was referring to.
пн, 17 сент. 2018 г. в 13:05, Maxim Muzafarov <maxmu...@gmail.com>: > Folks, > > Such warming up can be an effective technique for performing calculations > which required large cache > data reads, but I think it's the single narrow use case of all over Ignite > store usages. Like all other > powerfull techniques, we should use it wisely. In the general case, I think > we should consider other > techniques mentioned by Vladimir and may create something like `global > statistics of cache data usage` > to choose the best technique in each case. > > For instance, it's not obvious what would take longer: multi-block reads or > 50 single-block reads issues > sequentially. It strongly depends on used hardware under the hood and might > depend on workload system > resources (CPU-intensive calculations and I\O access) as well. But > `statistics` will help us to choose > the right way. > > > On Sun, 16 Sep 2018 at 23:59 Dmitriy Pavlov <dpavlov....@gmail.com> wrote: > > > Hi Alexei, > > > > I did not find any PRs associated with the ticket for check code changes > > behind this idea. Are there any PRs? > > > > If we create some forwards scan of pages, it should be a very > intellectual > > algorithm including a lot of parameters (how much RAM is free, how > probably > > we will need next page, etc). We had the private talk about such idea > some > > time ago. > > > > By my experience, Linux systems already do such forward reading of file > > data (for corresponding sequential flagged file descriptors), but some > > prefetching of data at the level of application may be useful for > O_DIRECT > > file descriptors. > > > > And one more concern from me is about selecting a right place in the > system > > to do such prefetch. > > > > Sincerely, > > Dmitriy Pavlov > > > > вс, 16 сент. 2018 г. в 19:54, Vladimir Ozerov <voze...@gridgain.com>: > > > > > HI Alex, > > > > > > This is good that you observed speedup. But I do not think this > solution > > > works for the product in general case. Amount of RAM is limited, and > > even a > > > single partition may need more space than RAM available. Moving a lot > of > > > pages to page memory for scan means that you evict a lot of other > pages, > > > what will ultimately lead to bad performance of subsequent queries and > > > defeat LRU algorithms, which are of great improtance for good database > > > performance. > > > > > > Database vendors choose another approach - skip BTrees, iterate > direclty > > > over data pages, read them in multi-block fashion, use separate scan > > buffer > > > to avoid excessive evictions of other hot pages. Corresponding ticket > for > > > SQL exists [1], but idea is common for all parts of the system, > requiring > > > scans. > > > > > > As far as proposed solution, it might be good idea to add special API > to > > > "warmup" partition with clear explanation of pros (fast scan after > > warmup) > > > and cons (slowdown of any other operations). But I think we should not > > make > > > this approach part of normal scans. > > > > > > Vladimir. > > > > > > [1] https://issues.apache.org/jira/browse/IGNITE-6057 > > > > > > > > > On Sun, Sep 16, 2018 at 6:44 PM Alexei Scherbakov < > > > alexey.scherbak...@gmail.com> wrote: > > > > > > > Igniters, > > > > > > > > My use case involves scenario where it's necessary to iterate over > > > > large(many TBs) persistent cache doing some calculation on read data. > > > > > > > > The basic solution is to iterate cache using ScanQuery. > > > > > > > > This turns out to be slow because iteration over cache involves a lot > > of > > > > random disk access for reading data pages referenced from leaf pages > by > > > > links. > > > > > > > > This is especially true when data is stored on disks with slow random > > > > access, like SAS disks. In my case on modern SAS disks array reading > > > speed > > > > was like several MB/sec while sequential read speed in perf test was > > > about > > > > GB/sec. > > > > > > > > I was able to fix the issue by using ScanQuery with explicit > partition > > > set > > > > and running simple warmup code before each partition scan. > > > > > > > > The code pins cold pages in memory in sequential order thus > eliminating > > > > random disk access. Speedup was like x100 magnitude. > > > > > > > > I suggest adding the improvement to the product's core by always > > > > sequentially preloading pages for all internal partition iterations > > > (cache > > > > iterators, scan queries, sql queries with scan plan) if partition is > > cold > > > > (low number of pinned pages). > > > > > > > > This also should speed up rebalancing from cold partitions. > > > > > > > > Ignite JIRA ticket [1] > > > > > > > > Thoughts ? > > > > > > > > [1] https://issues.apache.org/jira/browse/IGNITE-8873 > > > > > > > > -- > > > > > > > > Best regards, > > > > Alexei Scherbakov > > > > > > > > > > -- > -- > Maxim Muzafarov >