home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.protocols.nfs
- Path: sparky!uunet!gatech!asuvax!ennews!ennews.eas.asu.edu!mc
- From: mc@emma.la.asu.edu (Mike Caplinger)
- Subject: benchmarks for Sun NFS accelerators
- Message-ID: <MC.92Sep3105607@emma.la.asu.edu>
- Sender: news@ennews.eas.asu.edu (USENET News System)
- Organization: Mars Observer TES Project, ASU, Tempe AZ
- Date: Thu, 3 Sep 1992 17:56:07 GMT
- Lines: 92
-
- This message describes my recent benchmarks of NFS accelerator
- products for Sun SPARC computers. It is not an endorsement of any
- specific product, and like all benchmarks, is obviously dependent on
- the application. Most of our applications are image processing
- programs that read one or more files (some sequentially, some
- randomly) and write a result file, usually sequentially. These files
- are often tens or even hundreds of megabytes in size. We usually use
- mmap to map the files in their entirety into memory, both for input
- and output, and we always map shared so as to not use swap file space.
-
- Originally, we had intended to require local disk space on each
- workstation to store output files temporarily, because we knew NFS
- writing performance was poor. However, this approach is a pain to
- administrate (no one's local disk is ever big enough), so we wanted to
- see what various NFS accelerators would do for our applications.
-
- We wrote two simple benchmark programs to simulate the output phase of
- our applications. The first, called `write', does a single write
- system call of a specified amount of memory. The second, `map',
- creates, extends, and then mmaps a file of a given size, then
- sequentially touches the first byte of each page.
-
- Two products were tried: eNFS from Interstream and Sun's SBus
- Prestoserve (with 1 MB of NVRAM). We ran eNFS in both normal and
- asynchronous modes. We compare the performance results against local
- disk access; our goal was to find a solution that gave us performance
- within a factor of two of local access.
-
- The following results are for SBus Prestoserve versus local disk
- access. These numbers were obtained at the San Diego Sun office.
- Remote operations were from an IPX with 32 MB and a 669MB disk to a
- 630MP with 64MB(?), an attached 1.2GB Seagate disk, and two
- processors. Local operations were to the IPX's 669MB disk.
-
- file presto local presto local
- size(MB)write write map map
- 4 0:08 0:02 0:16 0:02
- 8 0:16 0:06 0:30 0:02
- 16 0:30 0:12 1:02 0:07
- 32 1:01 0:31 1:50 0:19
-
- The following results are for eNFS in normal mode. These numbers were
- obtained at MSSS. Remote operations were from an SS2 with 64 MB and
- an HP 1GB to a 670MP with 64MB, an attached 2GB Micropolis, and four
- processors. Local operations were to the SS2's HP disk.
-
- file eNFS local eNFS local
- size(MB)write write map map
- 4 0:15 0:03 0:49 0:01
- 8 0:30 0:05 1:39 0:01
- 16 1:06 0:10 3:22 0:03
- 32 2:33 0:29 6:35* 0:10
-
- * We sometimes saw faster times by a factor of two for this number,
- reason unknown, but it was never faster than about 3 minutes.
-
- These numbers, also obtained with the MSSS configuration, are for eNFS
- in asynchronous mode.
-
- file eNFS local eNFS local
- size(MB)write write map map
- 4 0:06 0:03 0:11 0:01
- 8 0:12 0:05 0:23 0:02
- 16 0:23 0:11 0:47 0:03
- 32 0:56 0:29 1:19 0:13
-
- This is a summary of the results for 32 MB output files, normalized to
- local performance:
-
- local presto eNFS eNFS async
- write 1 2.0 5.3 1.9
- map 1 5.8 39.5 6.1
-
- Disappointingly, neither eNFS nor Prestoserve comes close to meeting
- our requirements for the map benchmark. This is probably inevitable,
- given that the mapped pages are not written out to disk until the
- system runs out of memory if the mapped object is local, but they
- appear to be flushed when the process exits if the mapped object is
- remote. Once again, I wish that there was some way to specify the
- reliability semantics of NFS operations on a per-partition or even a
- per-file basis. mmap is superior for local access, but seems to
- overwhelm the available NFS accelerator products when pages are
- flushed to remote objects (perhaps an artifact of the way SunOS
- flushes pages?)
-
- Any comments or questions are welcome.
-
- Mike Caplinger, MSSS/Caltech Mars Observer Camera Project
- mc@msss.com
-
- disclaimer: this message contains the opinions of the author only, and
- is not the official position of MSSS or any of its funding agencies.
-