performance
I stumbled on this:
http://www.spec.org/osg/web99/results/res2000q4/
This shows some PA-RISC machines running HPUX performing reasonably well.
What would it take to get parisc-linux to perform similiarly to HPUX for these kinds of numbers? Showing that these kinds of numbers are possible for parisc-linux would certainly help in our quest to have real users.
- Alex
Signature · 4 lines
Alex deVries, Principal Solutions Architect, The Puffins at Linuxcare 613.562.2759 tel [email protected], http://www.linuxcare.com/ Linuxcare, Support for the revolution.
Quoted 4 lines · Alex deVries
I stumbled on this:
http://www.spec.org/osg/web99/results/res2000q4/This shows some PA-RISC machines running HPUX performing reasonably well.
HPUX11+Zeus on A500 posted the highest rated Specweb99 for single CPU machine. Though the HP specweb folks didn't expect that to last very long...
Note that IBM Netfinity server nearly doubled in performance using native OS+Zeus vs linux+Tux. I don't expect such a dramatic difference for parisc-linux+Tux vs HPUX+Zeus.
Quoted 4 lines
What would it take to get parisc-linux to perform similiarly to HPUX for these kinds of numbers? Showing that these kinds of numbers are possible for parisc-linux would certainly help in our quest to have real users.
Offhand, I can think of several things (roughly in order):
o Tux (and/or zero copy - aka "Block I/O")
o support more than 512MB memory - preferably 8 or 16GB (64-bit kernel)
o Profile Based Optimization (compiler/linker)
o Write a module to dump CPU performance counters to /proc (and use output to tune critical code paths)
o handle 10 packets per interrupt under acenic
o Jumbo Frames?
I'm sure others (eg jsm) could add alot more here.
grant
Grant Grundler
Unix Systems Enablement Lab
+1.408.447.7253
Quoted 4 lines
http://www.spec.org/osg/web99/results/res2000q4/
This shows some PA-RISC machines running HPUX performing reasonably well.
Specweb is basically a con. Its not a materially valuable benchmark. It benches how fast your irq handler path is, how fast your memory cache is, how good your DMA transfers are. It has almost nothing to do with how good a web server the box is.
Quoted 4 lines
What would it take to get parisc-linux to perform similiarly to HPUX for these kinds of numbers? Showing that these kinds of numbers are possible for parisc-linux would certainly help in our quest to have real users.
Basically it comes down to using DMA, caching everything and using the fact it is a completely hoax benchmark to write custom tuned code for the job. Even the people who 'use xyz webserver' and it sounds standard turn out to be embedding it or half of it in the kernel.
So it comes down to - tuning the network card driver, tuning the IRQ path, making sure the cache handling is optimal. The rest consists of loading the machine with 8Gbytes of RAM, using obscenely smart ethernet cards and having good memory bandwidth.
The 64bit pa-risc boxes have the hardware for it, in fact they have better hardware for it than x86. All you have to do is write better tlb, irq and pci handling code than the HP/UX engineering team 8).
Alan
Quoted 9 lines · Alan Cox
Quoted 4 lines
http://www.spec.org/osg/web99/results/res2000q4/
This shows some PA-RISC machines running HPUX performing reasonably well.
Specweb is basically a con. Its not a materially valuable benchmark. It benches how fast your irq handler path is, how fast your memory cache is, how good your DMA transfers are. It has almost nothing to do with how good a web server the box is.
I agree that specweb might not actually be functionally practical.
This might not be the case with specweb, but sometimes specifications do have the ability to influence customers regardless of their validity.
That's interesting for one reason: if the marketing helps make potential customers think that parisc-linux is fast, they'll buy parisc boxes, and HP is more likely to allocate developer time on this project.
- Alex
Signature · 4 lines
Alex deVries, Principal Solutions Architect, The Puffins at Linuxcare 613.562.2759 tel [email protected], http://www.linuxcare.com/ Linuxcare, Support for the revolution.
Alan,
Interesting comments. I had to think about them for a bit.
Quoted 4 lines · Alan Cox
Specweb is basically a con. Its not a materially valuable benchmark. It benches how fast your irq handler path is, how fast your memory cache is, how good your DMA transfers are. It has almost nothing to do with how good a web server the box is.
At some level, all benchmarks are bogus. Bogomips encapsulates the issue. Taking the benchmarks out of context is when the con game starts. Isn't Specweb just one aspect of what makes a good webserver? Others include cost, size, security, manageability, reliability, etc. It just seems CPU/Memory/IO "bandwidth" and utilization (efficiency) are valid things to measure. An outsiders perspective is Specweb attempts to encapsulate those numbers with a representative workload.
Marketing drives most benchmarking I've been involved in. And Marketing targets the more ignorant (but not too poor) customers. Smart (and rich) customers use specific benchmarks make a "short list" and then measure *their* application performance.
HP has several "Capacity Planning Centers" just for those customers.
[ deleted comment about putting webserver in kernel ]
Quoted 8 lines
So it comes down to - tuning the network card driver, tuning the IRQ path, making sure the cache handling is optimal. The rest consists of loading the machine with 8Gbytes of RAM, using obscenely smart ethernet cards and having good memory bandwidth.
The 64bit pa-risc boxes have the hardware for it, in fact they have better hardware for it than x86. All you have to do is write better tlb, irq and pci handling code than the HP/UX engineering team 8).
Given only a handful of people know the PARISC CPU/TLB as well as jsm, and I (re)wrote the HPUX PCI code, I suggest long term performance gains will be found in linux VM/PM (and other "architected" stuff like atomic operations - which will always suck on parisc) and drivers. (My definition of drivers includes networking stack and NIC device drivers). Short term goal is to optimize existing parisc-specific code around existing architecture.
[ /rant on ]
The HPUX engineering team is not lacking in talent - just opportunity to make the zillions of small changes required to improve scalability and reduce CPU utilization. Any single change resulting in less than 2% performance improvement on something like SpecWeb is very hard to get into HPUX kernel.
[ /rant off ]
grant
Quoted 7 lines
Given only a handful of people know the PARISC CPU/TLB as well as jsm, and I (re)wrote the HPUX PCI code, I suggest long term performance gains will be found in linux VM/PM (and other "architected" stuff like atomic operations - which will always suck on parisc) and drivers. (My definition of drivers includes networking stack and NIC device drivers). Short term goal is to optimize existing parisc-specific code around existing architecture.
Much of the work Ingo and others did concentrated on the eepro100 driver internals down to removing every pci stall, doing irq affinity and doing small optimisations on the networking stack.
So I'd say your suggestion matches what happened on x86 too. The atomic stuff is trickier. Linux is designed to optimal on the common high performance hardware and right now the design therefore focuses on ia32 at the cost sometimes of other systems.
Alan