[opensuse] Gobs of RAM?
Hi Folks, Does anyone have experience with openSUSE and lots of RAM? By lots, I mean as much as 6.1-TB of ECC RAM. Can Leap effectively use that much memory? Also, how about Python and matplotlib? I know there are better lists on which to ask that question, but the experience here runs deep and I bet someone knows. Or, can at least offer an opinion! :-) Thanks, Lew -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse+owner@opensuse.org
On Fri, Jun 3, 2016 at 1:15 PM, Lew Wolfgang
Hi Folks,
Does anyone have experience with openSUSE and lots of RAM? By lots, I mean as much as 6.1-TB of ECC RAM. Can Leap effectively use that much memory?
If I'm current on my CPU knowledge, you will need a quad CPU server MB with each core handling 1.5TB of the total RAM. Each quad CPU in turn will likely have 15 or more cores. So you're looking at roughly 60 cores (120 hyperthreads) That puts you into a serious NUMA design. The linux scheduler used in Leap 42.1 is known to have serious issues with NUMA efficiency https://blog.acolyer.org/2016/04/26/the-linux-scheduler-a-decade-of-wasted-c... Note that post is less than 2 months old and says: === In our experiments, these performance bugs caused many-fold performance degradation for synchronization-heavy scientific applications, 13% higher latency for kernel make, and a 14-23% decrease in TPC-H throughput for a widely used commercial database. === Given Leap 42.2 will be based on the already released 4.4 kernel, I doubt if Leap 42.2 will have all of that corrected. You will need to pay special attention to the Linux kernel you run and how many of the scheduler bugs identified in that report are fixed. Greg -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse+owner@opensuse.org
On 06/03/2016 10:41 AM, Greg Freemyer wrote:
On Fri, Jun 3, 2016 at 1:15 PM, Lew Wolfgang
wrote: Hi Folks,
Does anyone have experience with openSUSE and lots of RAM? By lots, I mean as much as 6.1-TB of ECC RAM. Can Leap effectively use that much memory?
If I'm current on my CPU knowledge, you will need a quad CPU server MB with each core handling 1.5TB of the total RAM. Each quad CPU in turn will likely have 15 or more cores. So you're looking at roughly 60 cores (120 hyperthreads)
That puts you into a serious NUMA design.
The linux scheduler used in Leap 42.1 is known to have serious issues with NUMA efficiency
https://blog.acolyer.org/2016/04/26/the-linux-scheduler-a-decade-of-wasted-c...
Note that post is less than 2 months old and says:
=== In our experiments, these performance bugs caused many-fold performance degradation for synchronization-heavy scientific applications, 13% higher latency for kernel make, and a 14-23% decrease in TPC-H throughput for a widely used commercial database. ===
Given Leap 42.2 will be based on the already released 4.4 kernel, I doubt if Leap 42.2 will have all of that corrected. You will need to pay special attention to the Linux kernel you run and how many of the scheduler bugs identified in that report are fixed.
Thanks for the info, Greg! In this case, I was looking at four Xeon E7-8893v3 CPU's each of which has four cores. I guess with hyperthreading that would give 32-cores. The homegrown Python/C code they run is single-threaded, so possibly scheduling wouldn't be a significant issue? The mobo would be a Supermicro 8048B-TR4FT. Interesting topic! Regards, Lew -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse+owner@opensuse.org
On Fri, Jun 3, 2016 at 2:14 PM, Lew Wolfgang
E7-8893v3
Per Intel: Recommended Customer Price $6841.00 So, over $25K just for the 4 CPUs Probably over $50K for the RAM. This is one hell of a machine you are talking about. All quad (or dual) socket servers are NUMA as I understand it. If you read the link I sent a big part of the issue is if one of the sockets runs out of work to do (ie. there aren't 8 threads scheduled to it), then the scheduler may fail to move new threads to it for seconds. That is even in the case of all the threads being independent. The Linux scheduler in the mainline kernel is simply not doing a good job of manager the NUMA thread pools. == next question == Is there much disk i/o? If so, NVMe SSDs are what you "want" for storage. The price is way down. A 128GB NVMe SSD from Samsung is under $100. (Scale up as needed.) NVMe solutions were in the thousands minimum a couple years ago. read these 2 posts about NVMe design from last night on the factory list: http://markmail.org/message/owyjv7di4ilss7bk http://markmail.org/message/wqpnm3yhyugm34ek == Highlevel data storage info == I've let myself get behind, but I think this is right. => SCSI /dev/sdx Been around forever, very feature rich. /dev/sdx originally meant SCSI disk x => IDE /dev/hdx The old ide drivers. Very popular 2000-2005. => SATA /dev/sdx Sata was IDE on steroids from a feature perspective. So much so, it was layered under the SCSI driver as libata instead of living in the old ide driver code. => NVMe /dev/nvme The newest thing under the sun. Designed explicitly to support highspeed SSDs and have direct connections to the PCIe bus. An entire new linux driver had to be developed. /dev/nvme0 is the 1st NVMe drive in your system, /dev/nvme0n1 is the 1st namespace (similar to a SCSI LUN) on that drive /dev/nvme0n1p1 is the 1st partition in that namespace. NVMe has a theoretical max speed of 10 GB/sec (iirc) Real world is above 1 GB/sec at affordable costs. I saw a laptop earlier this week with a 500GB NVMe SSD. (Made by Alienware). Greg -- Greg Freemyer www.IntelligentAvatar.net -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse+owner@opensuse.org
On 06/03/2016 12:04 PM, Greg Freemyer wrote:
On Fri, Jun 3, 2016 at 2:14 PM, Lew Wolfgang
wrote: E7-8893v3 Per Intel: Recommended Customer Price $6841.00
So, over $25K just for the 4 CPUs
Probably over $50K for the RAM.
This is one hell of a machine you are talking about.
The "pie-in-the-sky" quote was about $175,000 US. This included 400-GB of NVMe for the operating system. The data disks would be a RAID of spinning SAS. We've lots of experience here and can manage 1.4-GB/sec writing to an 11-disk RAID6 array. But it looks today like they choked on the price and are now considering modification of the display algorithms to run on smaller machines. Still, it would have been fun... Thanks for your help Greg, the vendor I was working with is going to look at the link you sent too. Regards, Lew -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse+owner@opensuse.org
On 06/03/2016 03:40 PM, Lew Wolfgang wrote:
On 06/03/2016 12:04 PM, Greg Freemyer wrote:
On Fri, Jun 3, 2016 at 2:14 PM, Lew Wolfgang
wrote: E7-8893v3 Per Intel: Recommended Customer Price $6841.00
So, over $25K just for the 4 CPUs
Probably over $50K for the RAM.
This is one hell of a machine you are talking about.
The "pie-in-the-sky" quote was about $175,000 US. This included 400-GB of NVMe for the operating system.
The data disks would be a RAID of spinning SAS. We've lots of experience here and can manage 1.4-GB/sec writing to an 11-disk RAID6 array.
But it looks today like they choked on the price and are now considering modification of the display algorithms to run on smaller machines. Still, it would have been fun...
Thanks for your help Greg, the vendor I was working with is going to look at the link you sent too.
Regards, Lew
Sheeze! And I thought I had seen a big box when I ended up with the hand-me-down Supermicro H8QM8-2. (the plus side to the older hardware was spare Opteron 8360 SE's were only $14 each, and an extra 32G of ram was another $40, I also splurged for a pair of LSI Megaraid 8888ELP cards $12 each [very little benefit in SATA III for platter drives] :-) Why anyone would ever need more than 64K of RAM?... Seriously, the limitations for the amount of memory that can be addressed/managed by each processor is ... well processor dependent. As for the Linux scheduler divvying out instructions into pipelines and the processor's scheduler and control of separate regions of RAM, there seems to be no one-size fits all solution. That's why you system designer folks make the big bucks :) That's about where my eyes glaze over. The discussions of NUMA addressing the processor stall problem inherent with SMP providing each core full access to all available memory makes sense, along with the NUMA nodes and memory zones, but then throw in the software dependency and differences needed to tailor different handling for small app small percentage of overall memory verses large application large percentage of ram use.... and I'll punt. A quick look and I found several articles that helped make it clear as mud: Non-uniform memory access https://en.wikipedia.org/wiki/Non-uniform_memory_access NUMA (Non-Uniform Memory Access): An Overview http://queue.acm.org/detail.cfm?id=2513149 I'm just glad I'm not the one that has to make the call on putting together six or seven figure boxes. I'll take comfort being the one that gets the call when something doesn't quite work out as expected :) -- David C. Rankin, J.D.,P.E. -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse+owner@opensuse.org
X-Originating-<%= hostname %>-IP: [88.97.62.77] On Fri, 2016-06-03 at 13:41 -0400, Greg Freemyer wrote:
On Fri, Jun 3, 2016 at 1:15 PM, Lew Wolfgang
wrote: [snip interesting and unusual question]
[snip interesting and informed reply with absorbing link] That's what I love about this list! Cheers, Dave -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse+owner@opensuse.org
03.06.2016 20:15, Lew Wolfgang пишет:
Hi Folks,
Does anyone have experience with openSUSE and lots of RAM? By lots, I mean as much as 6.1-TB of ECC RAM. Can Leap effectively use that much memory?
SLES can - there are in-memory database appliances with up to 15TB of memory, certified with SLES. Given that Leap kernel configuration should be close to SLES and - hopefully - should also include SLES patches I'd say "yes". Although of course *Leap* (or any operating system pure) does not need as much - the real question is, can your application sensibly use that amount of memory.
Also, how about Python and matplotlib? I know there are better lists on which to ask that question, but the experience here runs deep and I bet someone knows. Or, can at least offer an opinion! :-)
Thanks, Lew
-- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse+owner@opensuse.org
participants (5)
-
Andrei Borzenkov
-
Dave Howorth
-
David C. Rankin
-
Greg Freemyer
-
Lew Wolfgang