Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On 26/08/14 10:13, Josh Berkus wrote: On 08/22/2014 07:02 AM, Andres Freund wrote: On 2014-08-21 14:02:26 -0700, Josh Berkus wrote: On 08/20/2014 07:40 PM, Bruce Momjian wrote: Not sure how you can make such a blanket statement when so many people have tested and shown the benefits of hyper-threading. Actually, I don't know that anyone has posted the benefits of HT. Link? There's definitely cases where it can help. But it's highly workload *and* hardware dependent. The only cases I've seen where HT can be beneficial is when you have large numbers of idle connections. Then the idle connections can be "parked" on the HT virtual cores. However, even in this case I haven't seen a head-to-head performance comparison. I've just had a pair of Crucial m550's arrive, so a bit of benchmarking is in order. The results (below) seem to suggest that HT enabled is certainly not inhibiting scaling performance for single socket i7's. I performed several runs (typical results shown below). Intel i7-4770 3.4 Ghz, 16G 2x Crucial m550 Ubuntu 14.04 Postgres 9.4 beta2 logging_collector = on max_connections = 600 shared_buffers = 1GB wal_buffers = 32MB checkpoint_segments = 128 effective_cache_size = 10GB pgbench scale = 300 test duration (each) = 600s db on 1x m550 xlog on 1x m550 clients | tps (HT)| tps (no HT) +--+- 4 | 517 | 520 8 | 1013 | 999 16 | 1938 | 1913 32 | 3574 | 3560 64 | 5873 | 5412 128 | 8351 | 7450 256 | 9426 | 7840 512 | 9357 | 7288 Regards Mark -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On 26/08/14 10:13, Josh Berkus wrote: On 08/22/2014 07:02 AM, Andres Freund wrote: On 2014-08-21 14:02:26 -0700, Josh Berkus wrote: On 08/20/2014 07:40 PM, Bruce Momjian wrote: Not sure how you can make such a blanket statement when so many people have tested and shown the benefits of hyper-threading. Actually, I don't know that anyone has posted the benefits of HT. Link? There's definitely cases where it can help. But it's highly workload *and* hardware dependent. The only cases I've seen where HT can be beneficial is when you have large numbers of idle connections. Then the idle connections can be "parked" on the HT virtual cores. However, even in this case I haven't seen a head-to-head performance comparison. I recall HT beneficial on a single socket (i3 or i7), using pgbench as the measuring tool. However I didn't save the results at the time. I've just got some new ssd's to play with so might run some pgbench tests on my home machine (Haswell i7) with HT on and off. Regards Mark -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On 08/22/2014 07:02 AM, Andres Freund wrote: > On 2014-08-21 14:02:26 -0700, Josh Berkus wrote: >> On 08/20/2014 07:40 PM, Bruce Momjian wrote: >>> Not sure how you can make such a blanket statement when so many people >>> have tested and shown the benefits of hyper-threading. >> >> Actually, I don't know that anyone has posted the benefits of HT. >> Link? > > There's definitely cases where it can help. But it's highly workload > *and* hardware dependent. The only cases I've seen where HT can be beneficial is when you have large numbers of idle connections. Then the idle connections can be "parked" on the HT virtual cores. However, even in this case I haven't seen a head-to-head performance comparison. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On 2014-08-21 14:02:26 -0700, Josh Berkus wrote: > On 08/20/2014 07:40 PM, Bruce Momjian wrote: > > Not sure how you can make such a blanket statement when so many people > > have tested and shown the benefits of hyper-threading. > > Actually, I don't know that anyone has posted the benefits of HT. > Link? There's definitely cases where it can help. But it's highly workload *and* hardware dependent. > OS is RHEL with 2.6.32-431.3.1.el6.x86_64. > > I've emailed a kernel hacker who works at Intel for comment; for one > thing, I'm wondering if the older kernel version is a problem for a > system like this. I'm not sure if it has been backported by redhat, but there definitely have been significant improvement in SMT aware scheduling after vanilla 2.6.32. Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On 08/22/2014 01:37 AM, Scott Marlowe wrote: I thought they were fixed in 3.8.something? We're running 3.8 on our production servers but IO is not an issue for us. Yeah. 3.8 fixed a ton of issues that were plaguing us. There were still a couple patches I wanted that didn't get in until 3.11+, but the worst of the behavior was solved before that. Bugs in kernel cache page aging algorithms are bad, m'kay? -- Shaun Thomas OptionsHouse, LLC | 141 W. Jackson Blvd. | Suite 800 | Chicago IL, 60604 312-676-8870 stho...@optionshouse.com __ See http://www.peak6.com/email_disclaimer/ for terms and conditions related to this email -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On Thu, Aug 21, 2014 at 5:29 PM, Josh Berkus wrote: > On 08/21/2014 04:08 PM, Steve Crawford wrote: >> On 08/21/2014 03:51 PM, Josh Berkus wrote: >>> On 08/21/2014 02:26 PM, Scott Marlowe wrote: I'm running almost the exact same setup in production as a spare. It has 4 of those CPUs, 256G RAM, and is currently set to use HT. Since it's a spare node I might be able to do some testing on it as well. It's running a 3.2 kernel right now. I could probably get a later model kernel on it even. >>> You know about the IO performance issues with 3.2, yes? >>> >> Were those 3.2 only and since fixed or are there issues persisting in >> 3.2+? The 12.04 LTS release of Ubuntu Server was 3.2 but the 14.04 is 3.13. > > The issues I know of were fixed in 3.9. > I thought they were fixed in 3.8.something? We're running 3.8 on our production servers but IO is not an issue for us. -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On 08/21/2014 04:29 PM, Josh Berkus wrote: On 08/21/2014 04:08 PM, Steve Crawford wrote: On 08/21/2014 03:51 PM, Josh Berkus wrote: On 08/21/2014 02:26 PM, Scott Marlowe wrote: I'm running almost the exact same setup in production as a spare. It has 4 of those CPUs, 256G RAM, and is currently set to use HT. Since it's a spare node I might be able to do some testing on it as well. It's running a 3.2 kernel right now. I could probably get a later model kernel on it even. You know about the IO performance issues with 3.2, yes? Were those 3.2 only and since fixed or are there issues persisting in 3.2+? The 12.04 LTS release of Ubuntu Server was 3.2 but the 14.04 is 3.13. The issues I know of were fixed in 3.9. Correct. If you run trusty backports you are good to go. JD -- Command Prompt, Inc. - http://www.commandprompt.com/ 503-667-4564 PostgreSQL Support, Training, Professional Services and Development High Availability, Oracle Conversion, @cmdpromptinc "If we send our children to Caesar for their education, we should not be surprised when they come back as Romans." -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On 22/08/14 11:29, Josh Berkus wrote: On 08/21/2014 04:08 PM, Steve Crawford wrote: On 08/21/2014 03:51 PM, Josh Berkus wrote: On 08/21/2014 02:26 PM, Scott Marlowe wrote: I'm running almost the exact same setup in production as a spare. It has 4 of those CPUs, 256G RAM, and is currently set to use HT. Since it's a spare node I might be able to do some testing on it as well. It's running a 3.2 kernel right now. I could probably get a later model kernel on it even. You know about the IO performance issues with 3.2, yes? Were those 3.2 only and since fixed or are there issues persisting in 3.2+? The 12.04 LTS release of Ubuntu Server was 3.2 but the 14.04 is 3.13. The issues I know of were fixed in 3.9. There is a 3.11 kernel series for Ubuntu 12.04 Precise. Regards Mark -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On 08/21/2014 04:08 PM, Steve Crawford wrote: > On 08/21/2014 03:51 PM, Josh Berkus wrote: >> On 08/21/2014 02:26 PM, Scott Marlowe wrote: >>> I'm running almost the exact same setup in production as a spare. It >>> has 4 of those CPUs, 256G RAM, and is currently set to use HT. Since >>> it's a spare node I might be able to do some testing on it as well. >>> It's running a 3.2 kernel right now. I could probably get a later >>> model kernel on it even. >> You know about the IO performance issues with 3.2, yes? >> > Were those 3.2 only and since fixed or are there issues persisting in > 3.2+? The 12.04 LTS release of Ubuntu Server was 3.2 but the 14.04 is 3.13. The issues I know of were fixed in 3.9. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On 08/21/2014 03:51 PM, Josh Berkus wrote: On 08/21/2014 02:26 PM, Scott Marlowe wrote: I'm running almost the exact same setup in production as a spare. It has 4 of those CPUs, 256G RAM, and is currently set to use HT. Since it's a spare node I might be able to do some testing on it as well. It's running a 3.2 kernel right now. I could probably get a later model kernel on it even. You know about the IO performance issues with 3.2, yes? Were those 3.2 only and since fixed or are there issues persisting in 3.2+? The 12.04 LTS release of Ubuntu Server was 3.2 but the 14.04 is 3.13. Cheers, Steve -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On 08/21/2014 02:26 PM, Scott Marlowe wrote: > I'm running almost the exact same setup in production as a spare. It > has 4 of those CPUs, 256G RAM, and is currently set to use HT. Since > it's a spare node I might be able to do some testing on it as well. > It's running a 3.2 kernel right now. I could probably get a later > model kernel on it even. You know about the IO performance issues with 3.2, yes? -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
> HT off is common knowledge for better benchmarking result It's wise to use the qualifer 'for better benchmarking results'. It's worth keeping in mind here that a benchmark is not the same as normal production use. For example, where I work we do lots of long-running queries in parallel over a big range of datasets rather than many short-term transactions as fast as possible. Our biggest DB server is also used for GDAL work and R at the same time*. Pretty far from pgbench; not everyone is constrained by locks. I suppose that if your code is basically N copies of the same function, hyper-threading isn't likely to help much because it was introduced to allow different parts of the processor to be used in parallel when you're running hetarogenous code. But if you're hammering just one part of the CPU... well, adding another layer of logical complexity for your CPU to manage probably isn't going to do much good. Should HT be on or off when you're running 64 very mixed types of long-term queries which involve variously either heavy use of real number calculations or e.g. logic/string handling, and different data sets? It's a much more complex question than simply maxing out your pgbench scores. I don't have the data now unfortunately, but I remember seeing a benefit for HT on our 4 core e3 when running GDAL/Postgis work in parallel last year. It's not surprising though; the GDAL calls are almost certainly using different functions of the processor compared to postgres and there should be very little lock contention. In light of this interesting data I'm now leaning towards proposing HT off for our mapservers (which receive short, similar requests over and over), but for the hetaragenous servers, I think I'll keep it on for now. Graeme. * unrelated. There's also huge advantages for us in keeping these different programs running on the same machine since we found we can get much better transfer rates through unix sockets than with TCP over the network. -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On Thu, Aug 21, 2014 at 3:26 PM, Scott Marlowe wrote: > On Thu, Aug 21, 2014 at 3:02 PM, Josh Berkus wrote: >> On 08/20/2014 07:40 PM, Bruce Momjian wrote: >> >>> I am also >>> unclear exactly what you tested, as I didn't see it mentioned in the >>> email --- CPU type, CPU count, and operating system would be the minimal >>> information required. >> >> Ooops! I thought I'd posted that earlier, but I didn't. >> >> The processors in question is the Intel(R) Xeon(R) CPU E7- 4850, with 4 >> of them for a total of 40 cores or 80 HT cores. >> >> OS is RHEL with 2.6.32-431.3.1.el6.x86_64. > > I'm running almost the exact same setup in production as a spare. It > has 4 of those CPUs, 256G RAM, and is currently set to use HT. Since > it's a spare node I might be able to do some testing on it as well. > It's running a 3.2 kernel right now. I could probably get a later > model kernel on it even. > > -- > To understand recursion, one must first understand recursion. To update this last post, the machine I have is running ubuntu 12.04.1 right now, and I have kernels 3.2, 3.5, 3.8, 3.11, and 3.13 available to put on it. We're looking at removing it from our current production cluster so I could likely do all kinds of crazy tests on it. -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On Thu, Aug 21, 2014 at 3:02 PM, Josh Berkus wrote: > On 08/20/2014 07:40 PM, Bruce Momjian wrote: > >> I am also >> unclear exactly what you tested, as I didn't see it mentioned in the >> email --- CPU type, CPU count, and operating system would be the minimal >> information required. > > Ooops! I thought I'd posted that earlier, but I didn't. > > The processors in question is the Intel(R) Xeon(R) CPU E7- 4850, with 4 > of them for a total of 40 cores or 80 HT cores. > > OS is RHEL with 2.6.32-431.3.1.el6.x86_64. I'm running almost the exact same setup in production as a spare. It has 4 of those CPUs, 256G RAM, and is currently set to use HT. Since it's a spare node I might be able to do some testing on it as well. It's running a 3.2 kernel right now. I could probably get a later model kernel on it even. -- To understand recursion, one must first understand recursion. -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On Thu, Aug 21, 2014 at 02:17:13PM -0700, Josh Berkus wrote: > >> Actually, I don't know that anyone has posted the benefits of HT. Link? > >> I want to compare results so that we can figure out what's different > >> between my case and theirs. Also, it makes a big difference if there is > >> an advantage to turning HT on for some workloads. > > > > I had Greg Smith test my system when it was installed, tested it, and > > recommended hyper-threading. The system is Debian Squeeze > > (2.6.32-5-amd64), CPUs are dual Xeon E5620, 8 cores, 16 virtual cores. > > Can you post some numerical results? > > I'm serious. It's obviously easier for our users if we can blanket > recommend turning HT off; that's a LOT easier for them than "you might > want to turn HT off if these conditions ...". So I want to establish > that HT is a benefit sometimes if it is. > > I personally have never seen HT be a benefit. I've seen it be harmless > (most of the time) but never beneficial. I know that when hyperthreading was introduced that it was mostly a negative, but then this was improved, and it might have gotten bad again. I am afraid results are based on the type of CPU, so I am not sure we can know a general answer. I know I asked Greg Smith, and I assume he would know. -- Bruce Momjian http://momjian.us EnterpriseDB http://enterprisedb.com + Everyone has their own god. + -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On 08/21/2014 02:11 PM, Bruce Momjian wrote: > On Thu, Aug 21, 2014 at 02:02:26PM -0700, Josh Berkus wrote: >> On 08/20/2014 07:40 PM, Bruce Momjian wrote: >>> On Wed, Aug 20, 2014 at 12:13:50PM -0700, Josh Berkus wrote: On a read-write test, it's 10% faster with HT off as well. Further, from their production machine we've seen that having HT on causes the machine to slow down by 5X whenever you get more than 40 cores (as in 100% of real cores or 50% of HT cores) worth of activity. So we're definitely back to "If you're using PostgreSQL, turn off Hyperthreading". >>> >>> Not sure how you can make such a blanket statement when so many people >>> have tested and shown the benefits of hyper-threading. >> >> Actually, I don't know that anyone has posted the benefits of HT. Link? >> I want to compare results so that we can figure out what's different >> between my case and theirs. Also, it makes a big difference if there is >> an advantage to turning HT on for some workloads. > > I had Greg Smith test my system when it was installed, tested it, and > recommended hyper-threading. The system is Debian Squeeze > (2.6.32-5-amd64), CPUs are dual Xeon E5620, 8 cores, 16 virtual cores. Can you post some numerical results? I'm serious. It's obviously easier for our users if we can blanket recommend turning HT off; that's a LOT easier for them than "you might want to turn HT off if these conditions ...". So I want to establish that HT is a benefit sometimes if it is. I personally have never seen HT be a benefit. I've seen it be harmless (most of the time) but never beneficial. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On Thu, Aug 21, 2014 at 02:02:26PM -0700, Josh Berkus wrote: > On 08/20/2014 07:40 PM, Bruce Momjian wrote: > > On Wed, Aug 20, 2014 at 12:13:50PM -0700, Josh Berkus wrote: > >> On a read-write test, it's 10% faster with HT off as well. > >> > >> Further, from their production machine we've seen that having HT on > >> causes the machine to slow down by 5X whenever you get more than 40 > >> cores (as in 100% of real cores or 50% of HT cores) worth of activity. > >> > >> So we're definitely back to "If you're using PostgreSQL, turn off > >> Hyperthreading". > > > > Not sure how you can make such a blanket statement when so many people > > have tested and shown the benefits of hyper-threading. > > Actually, I don't know that anyone has posted the benefits of HT. Link? > I want to compare results so that we can figure out what's different > between my case and theirs. Also, it makes a big difference if there is > an advantage to turning HT on for some workloads. I had Greg Smith test my system when it was installed, tested it, and recommended hyper-threading. The system is Debian Squeeze (2.6.32-5-amd64), CPUs are dual Xeon E5620, 8 cores, 16 virtual cores. -- Bruce Momjian http://momjian.us EnterpriseDB http://enterprisedb.com + Everyone has their own god. + -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On 08/20/2014 07:40 PM, Bruce Momjian wrote: > On Wed, Aug 20, 2014 at 12:13:50PM -0700, Josh Berkus wrote: >> On a read-write test, it's 10% faster with HT off as well. >> >> Further, from their production machine we've seen that having HT on >> causes the machine to slow down by 5X whenever you get more than 40 >> cores (as in 100% of real cores or 50% of HT cores) worth of activity. >> >> So we're definitely back to "If you're using PostgreSQL, turn off >> Hyperthreading". > > Not sure how you can make such a blanket statement when so many people > have tested and shown the benefits of hyper-threading. Actually, I don't know that anyone has posted the benefits of HT. Link? I want to compare results so that we can figure out what's different between my case and theirs. Also, it makes a big difference if there is an advantage to turning HT on for some workloads. > I am also > unclear exactly what you tested, as I didn't see it mentioned in the > email --- CPU type, CPU count, and operating system would be the minimal > information required. Ooops! I thought I'd posted that earlier, but I didn't. The processors in question is the Intel(R) Xeon(R) CPU E7- 4850, with 4 of them for a total of 40 cores or 80 HT cores. OS is RHEL with 2.6.32-431.3.1.el6.x86_64. I've emailed a kernel hacker who works at Intel for comment; for one thing, I'm wondering if the older kernel version is a problem for a system like this. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On 08/20/2014 06:14 PM, Mark Kirkwood wrote: Notwithstanding the above results, my workmate Matt made an interesting observation: the scaling graph for (our) 60 core box (HT off), looks just like the one for our 32 core box with HT *on*. Hmm. I know this sounds stupid and unlikely, but has anyone actually tested PostgreSQL on a system with more than 64 legitimate cores? The work Robert Haas did to fix the CPU locking way back when showed significant improvements up to 64, but so far as I know, nobody really tested beyond that. I seem to remember similar choking effects when pre-9.2 systems encountered high CPU counts. I somehow doubt Intel would allow their HT architecture to regress so badly from Nehalem, which is almost 3-generations old at this point. This smells like something in the software stack, up to and including the Linux kernel. -- Shaun Thomas OptionsHouse, LLC | 141 W. Jackson Blvd. | Suite 800 | Chicago IL, 60604 312-676-8870 stho...@optionshouse.com __ See http://www.peak6.com/email_disclaimer/ for terms and conditions related to this email -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On 21/08/14 11:14, Mark Kirkwood wrote: You didn't mention what cpu this is for (or how many sockets etc), would be useful to know. Just to clarify - while you mentioned that the production system was 40 cores, it wasn't immediately obvious that the same system was the source of the measurements you posted...sorry if I'm being a mixture of pedantic and dense - just trying to make sure it is clear what systems/cpus etc we are talking about (with this in mind it never hurts to quote cpu and mobo model numbers)! Cheers Mark -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
> On Wed, Aug 20, 2014 at 12:13:50PM -0700, Josh Berkus wrote: >> On a read-write test, it's 10% faster with HT off as well. >> >> Further, from their production machine we've seen that having HT on >> causes the machine to slow down by 5X whenever you get more than 40 >> cores (as in 100% of real cores or 50% of HT cores) worth of activity. >> >> So we're definitely back to "If you're using PostgreSQL, turn off >> Hyperthreading". > > Not sure how you can make such a blanket statement when so many people > have tested and shown the benefits of hyper-threading. I am also > unclear exactly what you tested, as I didn't see it mentioned in the > email --- CPU type, CPU count, and operating system would be the minimal > information required. HT off is common knowledge for better benchmarking result, at least for me. I've never seen better result with HT on, except POWER. Best regards, -- Tatsuo Ishii SRA OSS, Inc. Japan English: http://www.sraoss.co.jp/index_en.php Japanese:http://www.sraoss.co.jp -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On Wed, Aug 20, 2014 at 12:13:50PM -0700, Josh Berkus wrote: > On a read-write test, it's 10% faster with HT off as well. > > Further, from their production machine we've seen that having HT on > causes the machine to slow down by 5X whenever you get more than 40 > cores (as in 100% of real cores or 50% of HT cores) worth of activity. > > So we're definitely back to "If you're using PostgreSQL, turn off > Hyperthreading". Not sure how you can make such a blanket statement when so many people have tested and shown the benefits of hyper-threading. I am also unclear exactly what you tested, as I didn't see it mentioned in the email --- CPU type, CPU count, and operating system would be the minimal information required. -- Bruce Momjian http://momjian.us EnterpriseDB http://enterprisedb.com + Everyone has their own god. + -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On Wed, Aug 20, 2014 at 1:36 PM, Shaun Thomas wrote: > That's so strange. Back when I did my Nehalem tests, we got a very strong > 30%+ increase by enabling HT. We only got a hit when we turned off turbo, or > forgot to disable power saving features. In my experience, it is crucially important to consider power saving features in most benchmarks these days, where that might not have been true a few years ago. The CPU scaling governor can alter the outcome of many benchmarks quite significantly. -- Regards, Peter Geoghegan -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On 21/08/14 07:13, Josh Berkus wrote: Mark, all: So, this is pretty damming: Read-only test with HT ON: [pgtest@db ~]$ pgbench -c 20 -j 4 -T 600 -S bench starting vacuum...end. transaction type: SELECT only scaling factor: 30 query mode: simple number of clients: 20 number of threads: 4 duration: 600 s number of transactions actually processed: 47167533 tps = 78612.471802 (including connections establishing) tps = 78614.604352 (excluding connections establishing) Read-only test with HT Off: [pgtest@db ~]$ pgbench -c 20 -j 4 -T 600 -S bench starting vacuum...end. transaction type: SELECT only scaling factor: 30 query mode: simple number of clients: 20 number of threads: 4 duration: 600 s number of transactions actually processed: 82457739 tps = 137429.508196 (including connections establishing) tps = 137432.893796 (excluding connections establishing) On a read-write test, it's 10% faster with HT off as well. Further, from their production machine we've seen that having HT on causes the machine to slow down by 5X whenever you get more than 40 cores (as in 100% of real cores or 50% of HT cores) worth of activity. So we're definitely back to "If you're using PostgreSQL, turn off Hyperthreading". Hmm - that is interesting - I don't think we compared read only scaling for hyperthreading on and off (only read write). You didn't mention what cpu this is for (or how many sockets etc), would be useful to know. Notwithstanding the above results, my workmate Matt made an interesting observation: the scaling graph for (our) 60 core box (HT off), looks just like the one for our 32 core box with HT *on*. We are wondering if a lot of the previous analysis of HT performance regressions should actually be reevaluated in the light of ...err is it just that we have a lot more cores...? [1] Regards Mark [1] Particularly as in *some* cases (single socket i7 for instance) HT on seems to scale fine. -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
On 08/20/2014 02:13 PM, Josh Berkus wrote: So we're definitely back to "If you're using PostgreSQL, turn off Hyperthreading". That's so strange. Back when I did my Nehalem tests, we got a very strong 30%+ increase by enabling HT. We only got a hit when we turned off turbo, or forgot to disable power saving features. -- Shaun Thomas OptionsHouse, LLC | 141 W. Jackson Blvd. | Suite 800 | Chicago IL, 60604 312-676-8870 stho...@optionshouse.com __ See http://www.peak6.com/email_disclaimer/ for terms and conditions related to this email -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
Re: [PERFORM] Turn off Hyperthreading! WAS: 60 core performance with 9.3
Mark, all: So, this is pretty damming: Read-only test with HT ON: [pgtest@db ~]$ pgbench -c 20 -j 4 -T 600 -S bench starting vacuum...end. transaction type: SELECT only scaling factor: 30 query mode: simple number of clients: 20 number of threads: 4 duration: 600 s number of transactions actually processed: 47167533 tps = 78612.471802 (including connections establishing) tps = 78614.604352 (excluding connections establishing) Read-only test with HT Off: [pgtest@db ~]$ pgbench -c 20 -j 4 -T 600 -S bench starting vacuum...end. transaction type: SELECT only scaling factor: 30 query mode: simple number of clients: 20 number of threads: 4 duration: 600 s number of transactions actually processed: 82457739 tps = 137429.508196 (including connections establishing) tps = 137432.893796 (excluding connections establishing) On a read-write test, it's 10% faster with HT off as well. Further, from their production machine we've seen that having HT on causes the machine to slow down by 5X whenever you get more than 40 cores (as in 100% of real cores or 50% of HT cores) worth of activity. So we're definitely back to "If you're using PostgreSQL, turn off Hyperthreading". -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance