Members of the LTC and OSDL(TM) have been working on a project
to study Linux (R) 2.5 performance on Hyperthreaded platforms.
We have results from running the OSDL-DBT2 OLTP workloads where
we compare 2.4.21rc1* to 2.5.68* and 2.5.68* with Ingo's
sched-2.5.68-B2 patch. We also investigated HT scaling by
comparing the metrics with 2.5.68* having hyperthreading turned
off.
We see exceptional improvement(64%)comparing 2.5.68 with HT on
versus HT off with the Cached CPU intensive workload. Surprisingly,
we also see very good improvement (22-23%) with the non-cached I/O
intensive workload.
This could be because we use the SAPDB parameter MAXCPU set to 8
matching the number of user task threads to the number of logical
CPUs in the HT case. Although there are many more threads created,
these 8 threads are very active. This scenario may represent the
case where CPU affinity and load balancing are at their optimal
situation.
We see very little difference between the 2.4 kernel and the 2.5
kernels having hyperthreading turned on for either workload. As we
understand, stock 2.4.21 doesn't have O(1) scheduler, stock 2.5.68
has O(1) scheduler, and stock 2.5.68+Ingo has hyperthreading patch ...
when the workload is subjected to those 2 schedulers, neither helps.
Any explanation?
Here are the details:
System:
Intel(R) Xeon(TM) 4 Physical CPUs CPUs @ 1.50GHz MHz System 4GB memory
Database: SAP DB 7.0.3.25
Cached DBT2 workload, (DATA_CACHE 300,000 8k pages)
Kernel__________2.4.21rc1 2.5.68 2.5.68Ingo 2.5.68(no HT)
Phys/Log Procs 8L/4P 8L/4P 8L/4P 4L/4P
Average 4184 4226 4226 2572.2
StdDev 46.98 26.95 17.45 146.09
Compare to
2.4.21rc1 (%) base(100) 101 101 60.87
Scaling rel to
2.5.68 no HT n/a 1.64 1.64 base(1.0)
Non Cached DBT2 workload, DATA_CACHE 300,000
Kernel__________2.4.21rc1 2.5.68 2.5.68Ingo 2.5.68(no HT)
Phys/Log Procs 8L/4P 8L/4P 8L/4P 4L/4P
Average 2256.25 2297.75 2274.25 1868.33
StdDev 2.22 10.05 22.74 19.01
Compare to
2.4.21rc1 (%) base(100) 101.84 100.8 81.31
Scaling rel to
2.5.68 no HT n/a 1.23 1.22 base(1.0)
Raw Data and vmstat plots, configs, readprofiles :
http://www.osdl.org/projects/dbt2dev/results/HT/68I/results.html
DBT2 workload information:
http://www.osdl.org/projects/performance/dbt2-page.html
*All kernels tested include a patch to eliminate bounce buffers
for the DAC960.
Regards,
Mary Edie Meredith (maryedie@osdl.org)
Mark Wong (mwong@osdl.org)
Open Source Development Lab
Duc Vianney (dvianney@us.ibm.com)
Linux Technology Center, IBM(R)
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Fri May 23 2003 - 22:00:45 EST