InterSystems Data Platforms and Performance Part 4 - Looking at Memory
This post will show you an approach to size shared memory requirements for database applications running on InterSystems data platforms including global and routine buffers, gmheap, and locksize as well as some performance tips you should consider when configuring servers and when virtualizing Caché applications. As ever when I talk about Caché I mean all the data platform (Ensemble, HealthShare, iKnow and Caché).
When I first started working with Caché most customer operating systems were 32-bit and memory for a Caché application was limited and expensive. Commonly deployed Intel servers had only a few cores and the only way to scale up was go with big iron servers or use ECP to scale out horizontally. Now even basic production grade servers have multiple processors, dozens of cores and minimum memory is 128 or 256 GB with the possibility of TB. For most database installations ECP is forgotten and we can now scale application transaction rates massively on a single server.
A key feature of Caché is the way we use data in shared memory usually referred to as database cache or global buffers. The short story is that if you can right size and allocate 'more' memory to global buffers you will usually improve system performance - data in memory is much faster to access than data on disk. Back in the day, when 32-bit systems ruled, the answer to the question how much memory should I allocate to global buffers? Was a simply - as much as possible! There wasn't that much available anyway, so sums were done diligently to calculate OS requirements, the number of and size of OS and Caché processes and real memory used by each to find the remainder to allocate as large a global buffer as possible.
The tide has turned
If you are running your application on a current generation server you can allocate huge amounts of memory to a Caché instance and a laissez-faire attitude often aplies because memory is now "cheap" and plentiful. However the tide has turned again and pretty much all but the very largest systems I see deployed now are virtualized. So while 'monster' VMs can have large memory footprints if needed the focus still comes back to right sizing systems. To make the most of server consolidation capacity planning is required that makes best use of available host memory.
What uses memory?
Generally there are four main consumers of memory on a Caché database server:
- Operating System, including filesystem cache.
- If installed, other non-Caché applications.
- Caché processes.
- Caché shared memory (includes global and routine buffers and GMHEAP).
At a high level the amount of physical memory required is to simply add up the requirements of each of the items on the list. All of the above use real memory, but they can also use virtual memory, a key part of capacity planning is to size a system so that there is enough physical memory so paging does not occur or is minimized, or at least minimize or eliminate hard page faults where memory has to be brought back from disk.
In this post I will focus on sizing Caché shared memory and some general rules for optimising memory performance. The operating system and kernel requirements vary by operating system, but will be several GB in most cases. File system cache varies and is will be whatever is available after the other items on the list take their allocation.
Caché is mostly processes - if you look at the operating system statistics while your application is running you will see cache processes (e.g. cache or cache.exe). So a simple way to observe what your application memory requirements are is to look at the operating system metrics. For example with
ps on Linux or
Windows process explorer and total the amount of real memory in use, extrapolating for growth and peak requirements. Be aware that some metrics report virtual memory which includes shared memory, so be careful to gather real memory requirements.
Sizing Global buffers - A simplified way
For a high transaction database one of the capacity planning goals is to size global buffers so that as much of the application database working set is in memory as possible. This will minimise read IOPS and generally make the application perform better. We also need to strike a balance so that other memory users such as operating system and Caché process are not paged out and and there is enough memory for filesystem cache.
I showed an example of what can happen if reads from disk are excessive in Part 2 of this series. In that case high reads were caused by a bad report or query, but the same effect can be seen if global buffers are too small forcing the application to be constantly reading data blocks from disk. As a sidebar its also worth noting that the landscape for storage is always changing - storage is getting faster and faster with advances in SSDs but data in memory close to the running processes is still best.
Of course every application is different so its important to say "your milage may vary", but there are some general rules which will get you started on the road to capacity planning shared memory for your application. After that you can tune for your specific requirements.
Where to start?
Unfortunately there is no magic answer... but as I talked about in the previous posts a good practice is to size systems CPU capacity so that for a required peak transaction rate CPU will be approximately 80% utilized at peak processing times. Leaving 20% headroom for short term growth or unexpected spikes in activity.
For example when I am sizing TrakCare systems I know CPU requirements for a known transaction rate from benchmarking and reviewing customer site metrics, and I can use a broad rule of thumb for Intel processor based servers:
Rule of thumb: Physical memory is sized at n GB per CPU core.
- For TrakCare database servers n is 8 GB. For smaller web servers its 4 GB.
Rule of thumb: Allocate n% of memory to Caché global buffers.
- For small to medium TrakCare systems n% is 60% leaving 40% of memory for operating system, filesystem cache and Caché processes. You may vary this, say to 50%, if you need a lot of filesystem cache or have a lot of processes. Or make it a higher % as you use very large memory configurations on large systems.
- This rule of thumb assumes only one Caché instance on the server.
So for example if the application needs 10 CPU cores then the VM would have 80 GB of memory, with 48 GB for global buffers and 32 GB for everything else.
Memory sizing rules apply to physical or virtualized systems, so for TrakCare VMs the same 1 vCPU : 8 GB memory ratio applies.
Tuning global buffers
There are a few items to observe to see how effective your sizing is. You can observe free memory outside Caché with operating system tools. Set up as per your best calculations then observe memory usage over time, and if there is always free memory the system can be reconfigured to increase global buffers or to right-size a VM.
Another key indicator of good global buffer sizing is to have read IOPS as low as possible - which means Caché cache efficiency will be high. You can observe the impact of different global buffer sizes on PhyRds and RdRatio with mgstat, an example of looking at these metrics is in Part 2 of this series. Unless you have your entire database in memory there will always be some reads from disk, the aim is simply to keep reads as low as possible.
Remember your hardware food groups and getting the balance right - more memory for global buffers will lower read IOPS but possibly increase CPU utilization because your system can now do more work in a shorter time - but lowering IOPS is pretty much always a good thing - your users will be happier with faster response times.
See the section below for applying your requirements to physical memory configuration.
For virtual servers plan not to oversubscribe your production VMs memory especially Caché shared memory, also more on this below.
Is your applications sweet spot 8GB of physical memory per CPU core? I can't say but see if a similar method works for your application. Whether its 4GB or 10GB per core. If you have found another method for sizing global buffers please leave a comment below.
Monitoring Global Buffer usage
The Caché utility
^GLOBUFF displays statistics about what your global buffers are doing at any point in time. For example to display the top 25 by percentage:
For example output could look like this:
Total buffers: 2560000 Buffers in use: 2559981 PPG buffers: 1121 (0.044%) Item Global Database Percentage (Count) 1 MyGlobal BUILD-MYDB1 29.283 (749651) 2 MyGlobal2 BUILD-MYDB2 23.925 (612478) 3 CacheTemp.xxData CACHETEMP 19.974 (511335) 4 RTx BUILD-MYDB2 10.364 (265309) 5 TMP.CachedObjectD CACHETEMP 2.268 (58073) 6 TMP CACHETEMP 2.152 (55102) 7 RFRED BUILD-RB 2.087 (53428) 8 PANOTFRED BUILD-MYDB2 1.993 (51024) 9 PAPi BUILD-MYDB2 1.770 (45310) 10 HIT BUILD-MYDB2 1.396 (35727) 11 AHOMER BUILD-MYDB1 1.287 (32946) 12 IN BUILD-DATA 0.803 (20550) 13 HIS BUILD-DATA 0.732 (18729) 14 FIRST BUILD-MYDB1 0.561 (14362) 15 GAMEi BUILD-DATA 0.264 (6748) 16 OF BUILD-DATA 0.161 (4111) 17 HISLast BUILD-FROGS 0.102 (2616) 18 %Season CACHE 0.101 (2588) 19 WooHoo BUILD-DATA 0.101 (2573) 20 BLAHi BUILD-GECKOS 0.091 (2329) 21 CTPCP BUILD-DATA 0.059 (1505) 22 BLAHi BUILD-DATA 0.049 (1259) 23 Unknown CACHETEMP 0.048 (1222) 24 COD BUILD-DATA 0.047 (1192) 25 TMP.CachedObjectI CACHETEMP 0.032 (808)
This could be useful in several ways, for example to see how much of your working set is kept in memory. If you find this utility is useful please make a comment below to enlighten other community users on why it helped you.
Sizing Routine Buffers
Routines your application is running including compiled classes are stored in routine buffers. The goal of sizing shared memory for routine buffers is for all your routine code to be loaded and stay resident in routine buffers. Like global buffers it is expensive and inefficient to read routines off disk. The maximum size of routine buffers is 1023 MB. As a rule you want more routine buffers than you need as there is always a big performance gain to have routines cached.
Routines buffers are made up of different sizes. By default Caché determines the number of buffers for each size, at install time the defaults for 2016.1 are 4, 16 and 64 KB. It is possible to change the allocation of memory for different sizes, however to start your capacity planning it is recommended to stay with Caché defaults unless you have a special reason for changing. For more information, see routines in the Caché documentation “config” appendix of the Caché Parameter File Reference and Memory and Startup Settings in the “Configuring Caché” chapter of the Caché System Administration Guide.
As your application runs routines are loaded off disk and stored in the smallest buffer the routine will fit. For example if a routine is 3 KB it will ideally be stored in a 4 KB buffer, if there are no 4 KB buffers available a larger one will be used. A routine larger than 32 KB will use as many 64 KB routine buffers as needed.
Checking Routine Buffer Use
mgstat metric RouLas
One way to understand if the routine buffer is large enough is the mgstat metric RouLas (routine loads and saves). A RouLas is a fetch from or save to disk. A high number of routine loads/saves may show up as a performance problem, in that case you can improve the performance by increasing the number of routine buffers.
If you have increased routine buffers to the maximum of 1023 MB and still find high RouLas a more detailed examination is available so you can see what routines are in buffers and how much is used with
ccontrol stat cache -R1
This will produce a listing of routine metrics including a list of routine buffers and all the routines in cache. For example a partial listing of a default Caché install is:
Number of rtn buf: 4 KB-> 9600, 16 KB-> 7200, 64 KB-> 2400, gmaxrouvec (cache rtns/proc): 4 KB-> 276, 16 KB-> 276, 64 KB-> 276, gmaxinitalrouvec: 4 KB-> 276, 16 KB-> 276, 64 KB-> 276, Dumping Routine Buffer Pool Currently Inuse hash buf size sys sfn inuse old type rcrc rtime rver rctentry rouname 22: 8937 4096 0 1 1 0 D 6adcb49e 56e34d34 53 dcc5d477 %CSP.UI.Portal.ECP.0 36: 9374 4096 0 1 1 0 M 5c384cae 56e34d88 13 908224b5 %SYSTEM.WorkMgr.1 37: 9375 4096 0 1 1 0 D a4d44485 56e34d88 22 91404e82 %SYSTEM.WorkMgr.0 44: 9455 4096 0 0 1 0 D 9976745d 56e34ca0 57 9699a880 SYS.Monitor.Health.x 2691:16802 16384 0 0 7 0 P da8d596f 56e34c80 27 383da785 START etc etc
"rtns/proc" on the 2nd line above is saying that 276 routines can be cached at each buffer size as default.
Using this information another approach to sizing routine buffers is to run your application and list the running routines with cstat -R1. You could then calculate the routine sizes in use, for example put this list in excel, sort by size and see exactly what routines are in use. If your are not using all buffers of each size then you have enough routine buffers, or if you are using all of each size then you need to increase routine buffers or can be more direct about configuring the number of each bucket size.
Lock table size
The locksiz configuration parameter is the size (in bytes) of memory allocated for managing locks for concurrency control to prevent different processes from changing a specific element of data at the same time. Internally, the in-memory lock table contains the current locks, along with information about the processes that hold those locks.
Since memory used to allocate locks is taken from GMHEAP, you cannot use more memory for locks than exists in GMHEAP. If you increase the size of locksiz, increase the size of GMHEAP to match as per the formula in the GMHEAP section below. Information about application use of the lock table can be monitored using the system management portal (SMP), or more directly with the API:
This API returns three values: "Available Space, Usable Space, Used Space". Check Usable space and Used Space to roughly calculate suitable values (some lock space is reserved for lock structure). Further information is available in Caché documentation.
Note: If you edit the locksiz setting, changes take place immediately.
The GMHEAP (the Generic Memory Heap) configuration parameter is defined as: Size (in kilobytes) of the generic memory heap for Caché. This is the allocation from which the Lock table, the NLS tables, and the PID table are also allocated.
Note: Changing GMHEAP requires a Caché restart.
To assist you in sizing for your application information about GMHEAP usage can be checked using the API:
This API also provides the ability to get available generic memory heap and recommended a GMHEAP parameters for configuration. For example, the
DisplayUsage method displays all memory used by each of the system components and the amount of available heap memory. Further information is available in Caché documentation.
To get an idea of GMHEAP usage and recommendations at any point in time you can use the
RecommendedSize method. However you will need to run this multiple times to build up a baseline and recommendations for your system.
Rule of thumb: Once again your applications mileage will vary, but somewhere to start your sizing could be one of the following:
(Minimum 128MB) or (64 MB * number of cores) or (2x locksiz) or whichever is larger.
Remember GMHEAP must be sized to include the lock table.
Mark Bolinsky wrote a great post explaining why turning on Huge pages in Linux is a great performance booster.
Danger! Windows Large Pages and Shared Memory
Caché uses shared memory on all platforms and versions and its a great performance booster, including on Windows where it is always used, but there are particular issues unique to Windows you need to be aware of.
When Caché starts, it allocates a single, large chunk of shared memory to be used for database cache (global buffers), routine cache (routine buffers), the shared memory heap, journal buffers, and other control structures. On Caché startup shared memory can be allocated using small or large pages. On Windows 2008 R2 and later Caché uses large pages by default, however if a system has been running for a long time, due to fragmentation contiguous memory may not be able to be allocated at Caché startup and Caché can instead start using small pages.
Unexpectedly starting Caché with small pages can cause Caché to start with less shared memory than defined in configuration or Caché may take a long time to start or fail to start. I have seen this happen on sites with a failover cluster where the backup server has not been used as a database server for a long time.
Tip: One mitigation strategy is to periodically reboot the offline Windows cluster server. Another is to use Linux.
Physical memory is dictated by the best configuration for the processor. A bad memory configuration can have a significant performance impact.
Intel Memory configuration best practice
This information applies to Intel processors only. Please confirm with vendors what rules apply to other processors.
Factors that determine optimal DIMMs performance include:
- DIMM type
- DIMM rank
- Clock speed
- Position to the processor (closest/furthest)
- Number of memory channels
- Desired redundancy features.
For example, on Nehalem and Westmere servers (Xeon 5500 and 5600) there are three memory channels per processor and memory should be installed in sets of three per processor. For current processors (for example E5-2600) there are four memory channels per processor, so memory should be installed in sets of four per processor.
When there are unbalanced memory configurations — where memory is not installed in sets of three/four or memory DIMMS are different sizes, unbalanced memory can impose a 23% memory performance penalty.
Remember that one of the features of Caché is in memory data processing so getting the best performance from memory is important. It is also worth noting that for maximum bandwidth servers should be configured for the fastest memory speed. For Xeon processors maximum memory performance is only supported at up to 2 DIMMs per channel, so the maximum memory configurations for common servers with 2 CPUs is dictated by factors including CPU frequency and DIMM size (8GB, 16GB, etc).
Rules of thumb:
- Use a balanced platform configuration: populate the same number of DIMMs for each channel and each socket
- Use identical DIMM types throughout the platform: same size, speed, and number of ranks.
- For physical servers round up total physical memory in a host server to the natural break points -- 64GB, 128GB and so on based on these Intel processor best practices.
VMware Virtualisation considerations
I will follow up in future with another post with more guidelines for when Caché is virtualized. However the following key rule should be considered for memory allocation:
Rule: Set VMware memory reservation on production systems.
As we have seen above when Caché starts, it allocates a single, large chunk of shared memory to be used for global and routine buffers, GMHEAP, journal buffers, and other control structures.
You want to avoid any swapping for shared memory so set your production database VMs memory reservation to at least the size of Caché shared memory plus memory for Caché processes and operating system and kernel services. If in doubt reserve the full production database VMs memory.
As a rule if you mix production and non-production servers on the same systems do not set memory reservations on non-production systems. Let non-production servers fight out whatever memory is left ;). VMware often calls VMs with more than 8 CPUs 'monster VMs'. High transaction Caché database servers are often monster VMs. There are other considerations for setting memory reservations on monster VMs, for example if a monster VM is to be migrated for maintenance or due to a High Availability triggered restart then the target host server must have sufficient free memory. There are stratagies to plan for this I will talk about them in a future post along with other memory considerations such as planning to make best use of NUMA.
This is a start to capacity planning memory, a messy area - certainly not as clear cut as sizing CPU. If you have any questions or observations please leave a comment.
As this entry is posted I am on my way to Global Summit 2016. If you are attending this year I will be talking about performance topics with two presentations, or I am happy to catch up with you in person in the developers area.