Thread

1. Performance problems on 4-way AMD Opteron 875 (dual core)

   Dirk Lutzebäck <lutzeb@aeccom.com> — 2005-08-05T11:11:31Z

   [[I'm posting this on behalf of my co-worker who cannot post to this 
   list at the moment]]
   
   Hi,
   
   I had installed PostgreSQL on a 4-way AMD Opteron 875 (dual core) and 
   the performance isn't on the expected level.
   
   Details:
   The "old" server is a 4-way XEON MP 3.0 GHz with 4MB L3 cache, 32 GB RAM 
   (PC1600)  and local FC-RAID 10. Hyper-Threading is off. (DL580)
   The "old" server is using Red Hat Enterprise Linux 3 Update 5.
   The "new" server is a 4-way Opteron 875 with 1 MB L2 cache, 32 GB RAM 
   (PC3200) and the same local FC-RAID 10. (HP DL585)
   The "new" server is using Red Hat Enterprise Linux 4 (with the latest 
   x86_64 kernel from Red Hat - 2.6.9-11.ELsmp #1 SMP Fri May 20 18:25:30 
   EDT 2005 x86_64)
   I use PostgreSQL version 8.0.3.
   
   The issue is that the Opteron is slower as the XEON MP under high load. 
   I have created a test with parallel queries which are typical for my 
   application. The queries are in a range of small queries (0.1 seconds) 
   and larger queries using join (15 seconds).
   The test starts parallel clients. Each clients runs the queries in a 
   random order. The test takes care that a client use always the same 
   random order to get valid results.
   
   Here are the number of queries which the server has finished in a fix 
   period of time.
   I used PostgreSQL 8.1 snapshot from last week compiled as 64bit binary 
   for DL585-64bit.
   I used PostgreSQL 8.0.3 compiled as 32bit binary for DL585-32bit and DL580.
   During the tests everything which is needed is in the file cache. I 
   didn't have read activity.
   Context switch  spikes are over 50000 during the test on both server. My 
   feeling is that the XEON has a tick more context switches.
   
   
   
   PostgreSQL params:
   max_locks_per_transaction = 256
   shared_buffers = 40000
   effective_cache_size = 3840000
   work_mem = 300000
   maintenance_work_mem = 512000
   wal_buffers = 32
   checkpoint_segments = 24
   
   
   I was expecting two times more queries on the DL585. The DL585 with 
   PostgreSQL 8.0.3 32bit does meltdown earlier as the XEON in production 
   use. Please compare 4 clients and 8 clients. With 4 clients the Opteron 
   is in front and with 8 clients the XEON doesn't meltdown that much as 
   the Opteron.
   
   I don't have any idea what cause this. Benchmarks like SAP's SD 2-tier 
   showing that the DL585 can handle nearly three times more load as the 
   DL580 with XEON 3.0. We choose the 4-way Opteron 875 based on such 
   benchmark to replace the 4-way XEON MP.
   
   Does anyone have comments or ideas on which I have to focus my work?
   
   I guess, the shared buffer cause the meltdown when to many clients are 
   accessing the same data.
   I didn't understand why the 4-way XEON MP 3.0 can deal with this better 
   as the 4-way Opteron 875.
   The system load on the Opteron is never over 3.0. The XEON MP has a load 
   up to 4.0.
   
   Should I try other settings for PostgreSQL in postgresql.conf?
   Should I try other setting for the compilation?
   
   I will compile the latest PostgreSQL 8.1 snapshot for 32bit to evaluate 
   the new shared buffer code from Tom.
   I think, the 64bit is slow because my queries are CPU intensive.
   
   Can someone provide a commercial support contact for this issue?
   
   Sven.
   
   

2. Re: Performance problems on 4-way AMD Opteron 875 (dual core)

   Michael Stone <mstone+postgres@mathom.us> — 2005-08-05T12:27:19Z

   On Fri, Aug 05, 2005 at 01:11:31PM +0200, Dirk Lutzebäck wrote:
   >I will compile the latest PostgreSQL 8.1 snapshot for 32bit to evaluate 
   >the new shared buffer code from Tom.
   >I think, the 64bit is slow because my queries are CPU intensive.
   
   Have you actually tried it or are you guessing? If you're guessing, then
   compile it as a 64 bit binary and benchmark that.
   
   Mike Stone
   
   

3. Re: Performance problems on 4-way AMD Opteron 875 (dual

   Dirk Lutzebäck <lutzeb@aeccom.com> — 2005-08-05T12:56:23Z

   Michael Stone wrote:
   
   > On Fri, Aug 05, 2005 at 01:11:31PM +0200, Dirk Lutzebäck wrote:
   >
   >> I will compile the latest PostgreSQL 8.1 snapshot for 32bit to 
   >> evaluate the new shared buffer code from Tom.
   >> I think, the 64bit is slow because my queries are CPU intensive.
   >
   >
   > Have you actually tried it or are you guessing? If you're guessing, then
   > compile it as a 64 bit binary and benchmark that.
   >
   > Mike Stone
   
   We tried it. 64bit 8.1dev was slower than 32bit 8.0.3.
   
   Dirk
   
   

4. Re: Performance problems on 4-way AMD Opteron 875 (dual core)

   Tom Lane <tgl@sss.pgh.pa.us> — 2005-08-05T13:53:14Z

   =?ISO-8859-1?Q?Dirk_Lutzeb=E4ck?= <lutzeb@aeccom.com> writes:
   > Here are the number of queries which the server has finished in a fix 
   > period of time.
   
   Uh, you never actually supplied any numbers (or much of any other
   specifics about what was tested, either).
   
   My first reaction is "don't vary more than one experimental parameter at
   a time".  There is no way to tell whether the discrepancy is due to the
   different hardware, different Postgres version, or 32-bit vs 64-bit
   build.
   
   			regards, tom lane