read(byte[] buf, int off, int len) has much better performance for large scans (e.g. while doing a merging compaction), if using 128KB buffers, it can make 80MB/s throughput, while read(long pos, byte[] buf, int off, int len) can only achieve 40MB/s. To make best utilization of HDFS performance, we should use positioned read for random read/small scans to achieve shorter response time. OTOH use plain read for large scans to have better throughput, and close-reopen the file after the scanner is destroyed to avoid socket congestion and fd leak. I then read the current CellCacheScanner code roughly and found it already has dealt with this issue! I guess Doug's already aware of this long before? Now I am all confused that we still have hundreds of tcp connections in FIN_WAIT1 state on every DataNode, I really wonder what's the real cause. Donald
On Sat, Mar 7, 2009 at 4:01 AM, Luke <[email protected]> wrote: > > Sorry, I didn't read your posts fully (was in a hurry). DfsBroker does > have pread interface, which is used by range server for random reads. > We just need to fix our HdfsBroker (PositionRead) to use HDFS > positioned read interface. Can you check if that improve things for > you? > > On Mar 6, 11:37 am, Luke <[email protected]> wrote: > > It appears that HDFS does have pread like interface: readFully(pos, > > buf, len). Can you run the tests again using this API and see if > > things improve? > > > > On Mar 6, 11:18 am, Luke Lu <[email protected]> wrote: > > > > > Great analysis Donald! Thanks for the numbers. It seems to me the > > > right fix would be enhance the HDFS client library to add a pread like > > > > interface to do the right thing for random reads. Maybe you want to > > > file a Hadoop jira ticket for that? > > > > > __Luke > > > > > On Mar 6, 2009, at 3:15 AM, Liu Kejia (Donald) wrote: > > > > > > On Thu, Mar 5, 2009 at 11:50 PM, donald <[email protected]> > wrote: > > > > > > So I have done more digging on this subject... > > > > > > There is another problem if many files are kept open at the same > time: > > > > once you read some data from a HDFS file by calling > FSInputStream.read > > > > (byte[] buf, int off, int len), a tcp connection between HdfsBroker > > > > and the DataNode that contains the file block is set up, this > > > > connection is kept until you read another block (by default 64MB in > > > > size) of the file, or close the file entirely. There is a timeout on > > > > the server side, but I see no clue on the client side. So you quickly > > > > end up with a lot of idle connections between the HdfsBroker and many > > > > DataNodes. > > > > > > What's even worse, no matter how many bytes the application wants to > > > > read, the HDFS client library always requests the the chosen DataNode > > > > to send all the remaining bytes of the block. Which means if you read > > > > 1 byte from the beginning of a block, the DataNode actually gets the > > > > request of sending the whole block, of which only the first few bytes > > > > are read. Consequences are: if the client reads nothing for quite a > > > > long while, 1) the kernel tcp send queue on the DataNode side and the > > > > tcp receive queue on the client side are quickly fed up; 2) the > > > > DataNode Xceiver thread (there is a max count of 256 by default) is > > > > blocked. Eventually the Xceiver timeouts, and closes the connection. > > > > However this FIN packet cannot reach client side as send&receive > > > > queues are still blocked. Here is what I observe from one node of our > > > > test cluster: > > > > $ netstat -ntp > > > > (Not all processes could be identified, non-owned process info > > > > will not be shown, you would have to be root to see it all.) > > > > Active Internet connections (w/o servers) > > > > Proto Recv-Q Send-Q Local Address Foreign > > > > Address State PID/Program name > > > > tcp 0 121937 10.65.25.150:50010 > > > > 10.65.25.150:38595 FIN_WAIT1 - > > > > [...] > > > > tcp 74672 0 10.65.25.150:38595 > > > > 10.65.25.150:50010 ESTABLISHED 32667/java > > > > [...] > > > > (and hundreds of other connections in the same states) > > > > > > Possible solutions without modifying hadoop client library are: 1) > > > > open-read-close the file stream every time the cell store is > accessed; > > > > 2) always use postioned read: read(long position, byte[] buf, int > off, > > > > int len) instead, because pread doesn't keep the tcp connection with > > > > DataNodes. Solution 1 is not scalable because every open() operation > > > > includes interaction with HDFS NameNode, which immediately becomes a > > > > bottleneck: in our test cluster the NameNode can only handle hundreds > > > > of parallel open() request per second, with an average delay of > 2-3ms. > > > > I haven't tested the performance of solution 2 yet, I will put up > some > > > > numbers tomorrow. > > > > > > Donald > > > > > > I've created 1000 files in a 11-node hadoop cluster, each file is > > > > 20MB. Then I wrote simple java programs to do the following tests: > > > > > > Opening all 1000 files, one process: about 2.5 s (2.5ms latency) > > > > Closing all 1000files, one process: 50ms > > > > Opening all 1000 files, 10 processes (running distributedly on the > > > > 10 datanodes): 15 s (15ms latency, or 700 opens/s) > > > > Reading the first 1KB data from each file (plain read), one process: > > > > > 6s (6ms latency) > > > > Reading the first 1KB data from each file (positioned read), one > > > > process: 2.5s (2.5ms latency) > > > > Reading the first 100KB data from each file, 1KB at a time > > > > (positioned read), one process: 130s (1.3ms latency, or 0.77MB/s) > > > > Reading the first 100KB data from each file, 1KB at a time (plain > > > > read), one process: 8.8s (11MB/s) > > > > > > The tests are done multiple times to make sure all blocks are > > > > effectively cached in Linux page cache.The hadoop version was 0.19.0 > > > > > with a few patches. io.file.buffer.size = 4096 > > > > > > Donald > > > > > > On Feb 25, 8:59 pm, "Liu Kejia (Donald)" <[email protected]> > wrote: > > > > > It turns out the hadoop-default.xml packaged in my custom > > > > > hadoop-0.19.0-core.jar has set the "io.file.buffer.size" to 131072 > > > > > (128KB), > > > > > which means DfsBroker has to open a 128KB buffer for every open > > > > file. The > > > > > official hadoop-0.19.0-core.jar sets this value to 4096, which is > > > > more > > > > > reasonable for applications like Hypertable. > > > > > Donald > > > > > > > On Fri, Feb 20, 2009 at 11:55 AM, Liu Kejia (Donald) > > > > > <[email protected]>wrote: > > > > > > > > Caching might not work very well because keys are randomly > > > > generated, > > > > > > resulting in bad locality... > > > > > > Even it's Java, hundreds of kilobytes per file object is still > > > > very big. > > > > > > I'll profile HdfsBroker to see what exactly is using so much > > > > memory, and > > > > > > post the results later. > > > > > > > > Donald > > > > > > > > On Fri, Feb 20, 2009 at 11:20 AM, Doug Judd > > > > <[email protected]> wrote: > > > > > > > >> Hi Donald, > > > > > > > >> Interesting. One possibility would be to have an open > > > > CellStore cache. > > > > > >> Frequently accessed CellStores would remain open, while seldom > > > > used ones get > > > > > >> closed. The effectiveness of this solution would depend on the > > > > > workload. > > > > > >> Do you think this might work for your use case? > > > > > > > >> - Doug > > > > > > > >> On Thu, Feb 19, 2009 at 7:09 PM, donald <[email protected]> > > > > > wrote: > > > > > > > >>> Hi all, > > > > > > > >>> I recently run into the problem that HdfsBroker throws out of > > > > memory > > > > > >>> exception, because too many CellStore files in HDFS are kept > > > > open - I > > > > > >>> have over 600 ranges per range server, with a maximum of 10 > cell > > > > > >>> stores per range, that'll be 6,000 open files at the same > > > > time, making > > > > > >>> HdfsBroker to take gigabytes of memory. > > > > > > > >>> If we open the CellStore file on demand, i.e. when a scanner is > > > > > >>> created on it, this problem is gone. However random-read > > > > performance > > > > > >>> may drop due to the the overhead of opening a file in HDFS. > > > > Any better > > > > > >>> solution? > > > > > > > >>> Donald > > > --~--~---------~--~----~------------~-------~--~----~ You received this message because you are subscribed to the Google Groups "Hypertable Development" group. To post to this group, send email to [email protected] To unsubscribe from this group, send email to [email protected] For more options, visit this group at http://groups.google.com/group/hypertable-dev?hl=en -~----------~----~----~----~------~----~------~--~---
