home *** CD-ROM | disk | FTP | other *** search
-
- I ran these tests about a week ago but couldn't get back to it until now.
- I've reduced the data and put this mail together quickly, so please forgive
- any typos.
-
- -chet
-
- **************************************************************
-
- I unpacked the tar file and tested the benchmark
- for sanity on abyss, a 4 processor 6000 system (6340 I believe).
- The system was pretty lightly loaded (20 users, load average 1.69).
-
- Here are the reduced results:
-
- Phase 1: 3.000000 seconds (min 3, max 3)
- Phase 2: 8.000000 seconds (min 8, max 8)
- Phase 3: 8.000000 seconds (min 8, max 8)
- Phase 4: 6.000000 seconds (min 6, max 6)
- Phase 5: 97.000000 seconds (min 97, max 97)
- Total: 122.000000 seconds
-
- So far, so good.
- Next, I connected a pair of DS5500s on a private net.
- Both systems are running ULTRIX V4.1.
-
- The server system has 32 MB, 25% bufcache, and RZ57 disks (the
- benchmark is loaded in the g partition of an RZ57).
-
- The client system has either 16 MB, 10% bufcache or
- 32 MB, 25% bufcache. The root filesystem (where /tmp is located)
- is the a partition of an RF71 disk (less performance than RZ57).
-
- I controlled for benchmark execution on server with benchmark
- directory local on server and /tmp local on server:
-
- + with and without presto
-
- I controlled for benchmark execution on client with benchmark
- directory remote on server and /tmp local on client:
-
- + with and without presto on server.
- + with and without presto on client.
- + 16 MB, 10% bufcache and 32 MB, 25% bufcache clients.
-
- With each configuration I ran three test iterations.
- Here are the reduced results:
-
- ===== Local Execution =====
-
- test1: 32 MB, 25% bufcache server
- execution on server, benchmark directory local, /tmp local
- no presto
-
- Phase 1: 2.666667 seconds (min 2, max 3)
- Phase 2: 3.333333 seconds (min 3, max 4)
- Phase 3: 6.333333 seconds (min 6, max 7)
- Phase 4: 5.666667 seconds (min 5, max 7)
- Phase 5: 60.000000 seconds (min 59, max 62)
- Total: 78.000000 seconds
-
- test2: 32 MB, 25% bufcache server
- execution on server, benchmark directory local, /tmp local
- presto
-
- Phase 1: 0.000000 seconds (min 0, max 0)
- Phase 2: 1.333333 seconds (min 1, max 2)
- Phase 3: 6.666667 seconds (min 6, max 8)
- Phase 4: 4.666667 seconds (min 4, max 5)
- Phase 5: 53.333332 seconds (min 53, max 54)
- Total: 66.000000 seconds
-
- ** From test2 we can see there is 12 seconds of disk head latency
- ** associated with synchronous writes in the benchmark across all
- ** filesystems.
-
- ===== Remote Execution, Small Client =====
-
- test3: 32 MB, 25% bufcache server
- 16 MB, 10% bufcache client
- execution on client, benchmark directory NFS, /tmp local
- no presto on server, no presto on client
-
- Phase 1: 3.000000 seconds (min 3, max 3)
- Phase 2: 11.333333 seconds (min 11, max 12)
- Phase 3: 13.000000 seconds (min 13, max 13)
- Phase 4: 8.000000 seconds (min 8, max 8)
- Phase 5: 82.666664 seconds (min 82, max 83)
- Total: 118.000000 seconds
-
- test4: 32 MB, 25% bufcache server
- 16 MB, 10% bufcache client
- execution on client, benchmark directory NFS, /tmp local
- presto on server, no presto on client
-
- Phase 1: 1.000000 seconds (min 1, max 1)
- Phase 2: 5.666667 seconds (min 5, max 6)
- Phase 3: 13.000000 seconds (min 13, max 13)
- Phase 4: 8.666667 seconds (min 8, max 9)
- Phase 5: 66.333336 seconds (min 66, max 67)
- Total: 94.666664 seconds
-
- ** Execution time reduced by 23 seconds (about 20%) with presto
- ** on server.
-
- test5: 32 MB, 25% bufcache server
- 16 MB, 10% bufcache client
- execution on client, benchmark directory NFS, /tmp local
- presto on server, presto on client
-
- Phase 1: 1.000000 seconds (min 1, max 1)
- Phase 2: 5.666667 seconds (min 5, max 6)
- Phase 3: 12.666667 seconds (min 12, max 13)
- Phase 4: 8.000000 seconds (min 8, max 8)
- Phase 5: 58.000000 seconds (min 58, max 58)
- Total: 85.333336 seconds
-
- ** From test5 we see there is about 9 seconds of disk head latency
- ** associated with synchronous writes in the benchmark across /tmp.
- ** Execution time reduced by about 28% with presto on both sides.
-
- ===== Remote Execution, Large Client =====
-
- test6: 32 MB, 25% bufcache server
- 32 MB, 25% bufcache client
- execution on client, benchmark directory NFS, /tmp local
- no presto on server, no presto on client
-
- Phase 1: 2.333333 seconds (min 2, max 3)
- Phase 2: 11.333333 seconds (min 11, max 12)
- Phase 3: 10.666667 seconds (min 10, max 11)
- Phase 4: 7.333333 seconds (min 7, max 8)
- Phase 5: 79.000000 seconds (min 79, max 79)
- Total: 110.666664 seconds
-
- ** 8 seconds reduced due to larger client cache compared to test3.
- ** Indeed, there is a reduction of 72% in read requests over the wire
- ** with the larger client cache.
-
- test7: 32 MB, 25% bufcache server
- 32 MB, 25% bufcache client
- execution on client, benchmark directory NFS, /tmp local
- presto on server, no presto on client
-
- Phase 1: 1.000000 seconds (min 1, max 1)
- Phase 2: 4.000000 seconds (min 4, max 4)
- Phase 3: 11.000000 seconds (min 11, max 11)
- Phase 4: 7.333333 seconds (min 7, max 8)
- Phase 5: 62.333332 seconds (min 62, max 63)
- Total: 85.666664 seconds
-
- ** Execution time reduced by 25 seconds (about 23%) with presto
- ** on server. This should be constant across runs if the volume of NFS
- ** write data is the same.
-
- test8: 32 MB, 25% bufcache server
- 32 MB, 25% bufcache client
- execution on client, benchmark directory NFS, /tmp local
- presto on server, presto on client
-
- Phase 1: 0.333333 seconds (min 0, max 1)
- Phase 2: 4.666667 seconds (min 4, max 5)
- Phase 3: 11.000000 seconds (min 11, max 11)
- Phase 4: 7.000000 seconds (min 7, max 7)
- Phase 5: 54.666668 seconds (min 54, max 55)
- Total: 77.666664 seconds
-
- ** From test7 we see there is about 8 seconds of disk head latency
- ** associated with synchronous writes in the benchmark across /tmp.
- ** This should be constant across runs if there is the same amount
- ** of local directory operations on the client.
- ** Execution time reduced by about 30% with presto on both sides.
- **
- **
- ** With presto on both sides, remote execution was equivalent
- ** to local execution without presto (test1).
-
-
-