Average CPU usage not fully utilized [closed] - windows-server-2008

This question is unlikely to help any future visitors; it is only relevant to a small geographic area, a specific moment in time, or an extraordinarily narrow situation that is not generally applicable to the worldwide audience of the internet. For help making this question more broadly applicable, visit the help center.
Closed 11 years ago.
We have created a multi-threaded application which process/parse big files (few hundred MB's) simultaneously. Application runs perfectly. But my client is disappointed the way cores of machine being used. He tried to watch the performance monitor and came to us with report. His point is if application is multi-threaded why CPU average utilization is below 25%. According to him, if nothing is running on system and file processing is taking time, CPU utilization should be more than 80-90%. I am not sure what answer or technical outcome will satisfy him. Please suggest.
EDIT
I have one multi-threaded application which loads the file from disc. After file is loaded in memory, i click on process button, and it starts parsing the file in memory. Lets assume for now, parsing is done in one thread. While app is parsing the data, my average CPU usage not fully used. What reason I can give to justify why CPU is not completely used. Any kind of report will do or technical documentation will help.

The question is very vague, but here are some general guidelines.
Disk IO is the main bottleneck unless the file processing is really fancy. Loading several files simultaneously will make this even worse, as the head needs to jump around (for non-SSD drives), and data will come into memory even slower. If you load four files at 100 MB each, that would already take around 4 seconds when done serially - and longer when done in parallel. Your program might or might not wait during that time and just don't process data at all.
So if your parallelization is mainly to process several files (one file per worker thread), then you might want to serialize the loading in one thread.
If you can work one file at a time, maybe your processing can be split up to work on different parts of the file, or the processing itself can take advantage of multiple CPUs (largely depends on your application).
If you need to write back data to disk - then this will be part of the game, too.
I think the main point here is minimizing IO delay (and a reasonable splitting of the workload between different CPUs).
Edit:
Of course take RAM into account - if you need to swap out, this will kill your performance instantly.
The best way is of course to go and profile...

Related

Abnormal query execution time [closed]

Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 7 years ago.
Improve this question
I'm execution the same query in two different environments
The first environment has 4GB RAM and an Intel processor 3,09 Ghz.
The second environment has 32GB RAM and an AMD processor 2,20 Ghz.
I'm wondering why the query is taking 6 minutes to be executed in the first environment and many many hours in the second one.
I checked the momory allocated for sql server by exectuting the query below, the two environments have the same value.
SELECT value_in_use
FROM sys.configurations
WHERE name = 'max server memory (MB)'
What could be the reason why the query is taking too much time in the second environment?
PS : The number of rows is the same in both environments (about 2 millions)
Several factors can make the results different. I hesitate to mention them because they are the kind of things people tend to reject, but I have experienced them all and had some confirmed by msft.
First, the person who suggested you look at execution plans is on the right track. That is likely to at least give you a clue as to what's different.
Reasons (assuming size and content of data are identical)
Statistics are different on the two servers, causing different execution plans.
Hardware performance is different. Slower or faster CPUs (different generation for example, even if clock speed is close), slower or faster disk. More cores vs. one core (leads to parallel plans vs. serial plans)
Hardware configuration is different (e.g. perormance difference). One uses a SAN one has direct storage. Sometimes matters a lot. Sometimes makes no difference once data is in cache.
Data storage is physically different. One server has greatly fragmented and sparse data blocks/pages in the table(s) or indexes you care about. One has compact and fast. This can occur due to different scenarios used to load data in the two systems.
SQLServer configuration settings: Limited Memory configuration on one system, for example.
Competing workload on one of the systems.
Missing or different indexes.
Different collation settings resulting in different index statistics and different plans.
Slightly different software version.
Size of data is about the same but content is different, changing query plan.
etc.
The time difference involved suggests a different execution plan is most likely.
If all else (indexes, etc.) is 'the same', sometimes it is worth forcing SQLServer to rerun statistics, then try the query again. Historically SQLServer has automatically updated statistics based on the volume of changes to data in a table. Sometimes this leads to statistics that are temporarily bad, as one table hits the threshold for an update but other tables do not, and the optimizer chooses bad plans. It all seems very improbable, but I once built a product that hit this problem regularly, and was fortunate enough to have a visiting MSFT SQLServer developer help me prove it.
Another that has really hit me is slow san or slow VMWare environment. Sans are often claimed to be fast, but turn out to deliver terrible performance to actual users due to slow network connections or competing workloads. VMWare or other virtualization environment often leads to the same problem, especially in large organizations that don't want to figure out what workload is hitting their clusters. Someone else's workload affects yours. So performance testing loses meaning.
My money is on some factor that changes the plan, like data content, statistics, index configuration.

Processing large amounts of files, slowed down by I/O, how to make more efficent?

I've been maintaining an application for a few years ever since creator stopped working on it. This application does batch operations on a large amount of pictures such as resizing, applying text, adding a logo, etc.
Lately I've been rewriting this application in order to improve efficiency and try to fully use every capability of the system by using all the available memory and splitting the operation over multiple threads.
The previous maintainer implemented multithreading in somewhat awful ways such as doing long operations in the UI thread (rendering the stop button he made inoperable) and using a really small thread pool.
So I've been optimizing code and trying out increasing the size of the thread pool but apparently nothing improves at all, the bottleneck seems to be reading and writing to disk.
Is there a way to improve this situation?
Determine exactly what IO the application does at the moment. Then try to convert it to sequential access patterns. Random IO is 100x slower on magnetic disks and 10x slower on SSDs.
If it is a magnetic disk make sure that only one thread writes sequentially at a time. If you're forced to do random IO tune the optimal queue length.

stress testing web applications on less capable hardware

My organization is having an interesting internal debate right now that raises a question that I would like to open to the community at large.
The issue at hand is our environment in which we do stress-testing, capacity-testing, performance-regression-testing, and the like.
On one side of the debate are some software engineers who would like this environment to mirror the production environment as much as possible, in the interest of making the results as meaningful as possible. While we currently do have an environment for such testing, it is far less capable than the production system, and these software engineers feel that they are reaching the limits of what they can learn from it.
On the other side of the debate are some network engineers who both administer the environments and control the purse-strings. While they concede that capacity-testing would be better in an environment that is a better replica of the production environment, they argue that – for the purposes of stress testing – a more modest environment would have the effect of magnifying performance bottlenecks, making them easier to discover and fix.
This finally brings us to the part that piqued my interest: one software engineer suggests that while a more modest stress environment will increase the likelihood that you will encounter some bottleneck, it does not necessarily follow that it would help you find the next bottleneck you may encounter in production. The scaling effect, he argues, may not be linear.
Is there merit to that point of view? If yes, then why? What are the sources of that nonlinearity?
There are a lot of moving parts involved here: a cluster of java application servers, a cluster of database servers, lots of dynamic content being generated for each HTTP hit.
Edit: I appreciate everybody's thoughts so far, but I was really hoping that someone would do more than re-affirm one side or the other and actually tackle the question of "why". If there is such nonlinearity, what gives rise to it? Better yet it would be great if the reasons were expressed in terms of the CPU, memory, bandwidth, latency, interactions between subsystems, what have you... TerryE, you have come the closest. You should re-post your comment as an answer for the bounty if no one else steps up
Your software developer is right and I will take the point even further.
When you test an application components, like a web service, to see its behaviour under load, it is understandable to use a less capable environment. You can find the bottlenecks about memory, io etc. And most probably will find bugs and oversights like out of memory errors and log files getting huge.
But when your application components are running as intended and you need to test the whole shebang, you need to test the real environment.
When you run stress tests on an environment, you measure that environment's behaviour under load and its bottlenecks. While this tests may provide valuable information, this information will not be about your production system. The bottlenecks you find might not be relevant to your real system and you may spend precious development time to fix the bugs that do not exist. To know about bottlenecks you really might face with, you should run your stress tests on your real production system (preferably before the grand opening).
The assumption of the network engineers is that modest system is basically a scale model of the production system. They are also assuming that the various characteristics of the production environment which would be affecting the software performance are mirrored in the more modest system just at lower levels however in the same ratios. For instance, the CPU is not as fast, there is not quite as much memory, the storage is a bit slower, etc. and all of these differences are in similar ratios such that if everything were magically multiplied by some factor, say 1.77, the resulting changed modest system would be exactly like the production system.
However that the modest system is an exact scale model in all particulars of the production system is difficult for me to believe.
Here is a specific example. Lets say that measurements on the production system indicates that CPU utilization, the percentage of time the CPU is not idle, is too high. So you put the software on the modest system and do measurements and discover that on the modest system, the CPU utilization is lower. An investigation reveals that the modest system has slower storage so the CPU is spending more time idle waiting on data transfer from storage to complete because the application is I/O bound on the modest system where as on the production system it is not. This difference is due to the modest machine not being an exact scale model of the production machine because the CPU ratio is different than the I/O transfer ratio.
Another example would be having more memory allowing fewer page faults in the production environment. When the software is loaded onto the more modest machine, there are more page faults due to having less physical memory. With the various applications paging in and out, they begin to affect each other as pages of other applications are swapped out and then swapped back in again. On the product machine with larger memory, this cascading page fault behavior is not seen because there is sufficient memory to hold more applications simultaneously.
The point that I am really trying to make here is that a computer with all its various parts and applications is a complex, dynamic system. The idea that one computing environment is just a scale model of another is too simplistic of an assumption. Using a modest system can certainly provide valuable data. However once the gross adjustments have been made to the software and you are beginning to get into more subtle detailed adjustments, the differences in the environment can have a large impact on the results of the testing.
Some citations.
Computer systems are dynamical systems by Tod Mytkowicz, Amer Diwan, and Elizabeth Bradley.
Bayesian fault detection and diagnosis in dynamic systems by Uri Lerner, Ronald Parr, Daphne Koller, and Gautam Biswas.
I have encountered the similar situation in my production environment. We use modest system just for initial and basic level testing and findings. It is true that you can never find real bottlenecks and other performance issues on your testing environment. So to find real performance related issues and to find bottlenecks you must do it on production environment, there's no other way.
We have hosted over 2.5Million websites, although it might not be case with yours but let me tell you this, that in our case, we have faced horrible situations of linear bottlenecks. Meaning, we first faced memory issues when our traffic was getting increased. We resolved that by adding more memory. Until then we didn't even notice that having only 256 threads of httpd was our next bottleneck because limited memory was hiding it, once we resolved memory issue it quickly came down to the problem that why our webservers were slow again after just few weeks? We found out that 256 httpd threads are just not enough to serve that much traffic. We not only increased threads but also installed HA parallel load balancers in front of our WebServers to mitigate the issue.
Fortunately it solved our slow page loading problems. But after few months as traffic continuously grew we got into next bottleneck of storage system. You know what this time disk I/Os was the issue. To make the story short, we put parallel NFS based physical storage systems. Each NFS machine now serve files by having over 2000 threads running.
I forgot to mentioned that Database was also a big culprit of slowness that we resolved that issue by installing Master-Slaves model in cluster. We had to do a lot of performance tweaks in our application code as well and we had to physically distribute our application into different modules over different servers.
I'm just mentioning all this to prove a point that it's very likely that all performance related issues almost come in a linear way, at least that's what we have faced in our WebBased model. Even you have tested a lot on your modest systems you still have chances hidden bottlenecks which you can't find on testing environments.
What I have learned in my last 6 years experience that try to DISTRIBUTE your model AS MUCH AS POSSIBLE if you think you might going to have a lot of traffic or hits/sec. Centralized model can hold your traffic for some time by doing much tweaks but in the end your system gets busted.
I'm not saying you will face some bottlenecks or issues in your situation but I just wanted to warn you that these cases happen sometimes, just so you better aware.
**Sorry for my English.
good question. learning and optimization is best on modest hardware. but testing is safer on mirror (or at least something from same epoch)
it seams like you try to predict the first bottleneck that will appears and when it will happen. i'm not sure if that's the correct objective and the correct way. i assume we don't speak about a typical CRUD where client says 'it should work as fast as every other web application'. if you want to do tests correctly then, before you start your tests, you should know the expected load. expected number of users, expected number of events, response time etc. it's a part of your product specification. if you don't have the numbers, that means your analysts didn't do their job.
if you have the numbers then you don't need exact tests result. you just need to know the order of magnitude. you should also check how your software/hardware scale. how many instances do you need to handle x users/requests/whatever and how many to handle y
We load test systems for our customers every day -- and we see a wide range of problems. Certain classes of problems can be found on down-sized systems. Other cannot. Some can ONLY be found in production...because no matter how closely you mirror the two systems, they can never be identical. You can get REALLY close, if you work hard enough.
So, simple fact of testing: the closer your system is to the production system, the more accurate your tests will be.
IMO, this is one of the best reasons for moving to the cloud: you can spin up a system that is very close to your production system (about as identical as you could ever get) and run your load tests on that.
It is probably worth mentioning that we've occasionally seen customers waste a lot of hours chasing problems in their test environments that never would have occurred in production. The more different the environments are, the more likely this is to happen :(
I think you have partially answered your own question - you already have a production level environment and are already finding it is not at the same level / not as capable as the production environment. The bottom line is that with all the money in the world you will never be able to replicate the exact functioning of the production website - timings of events, volumes, cpu utilisation, memory utilisation, db IO, when it's all working in anger the behaviour can be non-deterministic to a certain extend. My point is you can never make it exactly the same. And on the other side of the coin a production environment by it's nature is going to be an expensive environment with a lot of kit in order to make it perform and handle your production volume of data / transactions. This is a big expense / overhead to the business, and in these times of frugality should we not be looking to avoid additional cost to the business.
Maybe a different tactic should be taken - learn the performance profile of your production software - how it scales with volume, does running times increase linearly, exponentially or logarithmically? Can you model this? Firstly you can verify that the test environment is behaving in a similar way - this is key to having a valid test. Then the other important part is taking relative tests rather than absolutes - you aren't going to get absolute running times that are the same as production, but run your performance tests before deploying the code changes to give you your baseline, then deploy your code changes and re-run the performance tests - this will give you the relative changes in production (e.g. will the performance degrade with this code release), based on your models of performance you will be able to verify that the software is scaling in the same way with extra volume.
So my viewpoint is that there is a great deal you can learn about your software and hardware performance in the lower environment, and doing this on a smaller / less capable infrastructure saves your company money, and if used right can give you most of your answers to performance testing that you are looking for.

Improve MPI program

I wrote a MPI program that seems to run ok, but I wonder about performance. Master thread needs to do 10 or more times MPI_Send, and the worker receives data 10 or more times and sends it. I wonder if it gives a performance penalty and whether I could transfer everything in single structs or which other technique could I benefit from.
Other general question, once a mpi program works more or less, what are the best optimization techniques.
It's usually the case that sending 1 large message is faster than sending 10 small messages. The time cost of sending a message is well modelled by considering a latency (how long it would take to send an empty message, which is non-zero because of the overhead of function calls, network latency, etc) and a bandwidth (how much longer it takes to send an extra byte given that the network communications has already started). By bundling up messages into one message, you only incurr the latency cost once, and this is often a win (although it's always possible to come up with cases where it isn't). The best way to know for any particular code is simply to try. Note that MPI datatypes allow you very powerful ways to describe the layout of your data in memory so that you can take it almost directly from memory to the network without having to do an intermediate copy into some buffer (so-called "marshalling" of the data).
As to more general optimization questions about MPI -- without knowing more, all we can do is give you advice which is so general as to not be very useful. Minimize the amount of communications which need to be done; wherever possible, use built-in MPI tools (collectives, etc) rather than implementing your own.
One way to fully understand the performance of your MPI application is to run it within the SimGrid platform simulator. The tooling and models provided are sufficient to get realistic timing predictions of mid-range applications (like, a few dozen thousands lines of C or Fortran), and it can be associated to adapted visualization tools that can help you fully understand what is going on in your application, and the actual performance tradeoffs that you have to consider.
For a demo, please refer to this screencast: https://www.youtube.com/watch?v=NOxFOR_t3xI

Scattered-write speed versus scattered-read speed on modern Intel or AMD CPUs?

I'm thinking of optimizing a program via taking a linear array and writing each element to a arbitrary location (random-like from the perspective of the CPU) in another array. I am only doing simple writes and not reading the elements back.
I understand that a scatted read for a classical CPU can be quite slow as each access will cause a cache miss and thus a processor wait. But I was thinking that a scattered write could technically be fast because the processor isn't waiting for a result, thus it may not have to wait for the transaction to complete.
I am unfortunately unfamiliar with all the details of the classical CPU memory architecture and thus there may be some complications that may cause this also to be quite slow.
Has anyone tried this?
(I should say that I am trying to invert a problem I have. I currently have an linear array from which I am read arbitrary values -- a scattered read -- and it is incredibly slow because of all the cache misses. My thoughts are that I can invert this operation into a scattered write for a significant speed benefit.)
In general you pay a high penalty for scattered writes to addresses which are not already in cache, since you have to load and store an entire cache line for each write, hence FSB and DRAM bandwidth requirements will be much higher than for sequential writes. And of course you'll incur a cache miss on every write (a couple of hundred cycles typically on modern CPUs), and there will be no help from any automatic prefetch mechanism.
I must admit, this sounds kind of hardcore. But I take the risk and answer anyway.
Is it possible to divide the input array into pages, and read/scan each page multiple times. Every pass through the page, you only process (or output) the data that belongs in a limited amount of pages. This way you only get cache-misses at the start of each input page loop.