From: Iozone [mailto:capps@xxxxxxxxxx] Sent: Wednesday, June 04, 2014 11:39 AM To: linux-nfs@xxxxxxxxxxxxxxx Subject: Forwarding request at suggestion from support Dear kernel folks, Please take a look at Bugzilla bug: https://bugzilla.redhat.com/show_bug.cgi?id=1104696 Description of problem: Linux NFSv3 clients can issue extra reads beyond EOF. Condition of the test: (32KB_file is a file that is 32KB in size) File is being read over an NFSv3 mount. dd if=/mnt/32KB_file of=/dev/null iflag=direct bs=1M count=1 What one should expect over the wire: NFSv3_read for 32k, or NFS_read for 1M NFSv3_read Reply return of 32KB and EOF set. What happens with Linux NFSv3 client: NFSv3 read for 128k NFSv3 read for 128k, NFSv3 read for 128k, NFSv3 read for 128k, NFSv3 read for 128k, NFSv3 read for 128k, NFSv3 read for 128k, NFSv3 read for 128k. followed by: NFSv3 read reply of 32k, NFSv3 read reply of 0, NFSv3 read reply of 0, NFSv3 read reply of 0, NFSv3 read reply of 0, NFSv3 read reply of 0, NFSv3 read reply of 0, NFSv3 read reply of 0. So? instead of a single round trip with a short read length returned, there were 8 async I/O ops sent to the NFS server, and 8 replies from the NFS server. The client knew the file size before even sending the very first request, but went ahead and issued an large number of reads that it should have known were beyond EOF. This client behavior hammers NFS servers with requests that are guaranteed to always fail, and burn CPU cycles, for operations that it knew were pointless. While the application is getting correct answers to the API calls, the poor client and server are beating each other senseless over the wire. NOTE: This only happens if O_DIRECT is being used? (thus the iflag=direct) Version-Release number of selected component (if applicable): Every version of Linux that I have tested seems to do this... How reproducible: Extremely. Though the actual transfer size being sent by the Linux client is sometimes a bunch of 32k transfers, or 128k or 256k, depending on the version of the kernel, and the NFS mount block size. A worse case is app does 1MB read (with O_DIRECT), with NFS block size set to 32K, the number of async reads is 1 useful and 31 that were beyond EOF. So for each 32k of file data, there are 62 extra NFS messages between the client and the server, and only 2 messages that made sense or ever should have taken place. ( IMHO ) Steps to Reproduce: The steps are above. In the attachment, the file size is 32k, the NFS block size is 32k. So you can see all of the extra async (back to back) client requests that are all going to return zero, except the very first one. ---------------------------------------------------------------------------- ----------------------------------- More details of why this is an issue: In the SPECsfs2014 benchmark (under development) there is a workload that simulates a software build environment. There are billions of small files. One of the operations that is tested is to read an entire file. This operation uses a large transfer size so that it can read the files as efficiently as possible. Whenever the file size is smaller than this large transfer size, the NFS client issues 8 to 64 times as many I/Os as were necessary to read the small file. If you take into consideration that this benchmark is simulating hundreds, or thousands of users, with billions of files, and is using multiple client nodes to present load to the server under test?. This overshoot on the reads is burying the NFS server with work that should never have been sent to the server. Instead of the benchmark measuring how fast the NFS server can serve files, it becomes a test of how many insane requests beyond EOF can the server tolerate from a Linux NFSv3 client, while serving almost no file data at all. I don?t understand why the Linux NFS client would ever issue reads beyond EOF. These files were opened, (LOOKUP, GETATTR, ACCESS) so the Linux kernel knows the file size. It should be a one line change to the code to simply *not* issue async read-aheads for file data that is beyond EOF. AND?. The Linux NFS client code certainly appears to know how to not read beyond EOF when the O_DIRECT flag is off. Why is this not the same when the O_DIRECT flag is on ? Thank you, Don Capps Capps at iozone dot org P.S. This overshoot is also confirmed to be happening in the NFSv4 client code. -- To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
