Linux’ kscand

In Linux the kscand process is a kernel process which is a key component of the virtual memory (vm) system.  According to Unix Tech Tips & Tricks’ excellent Understanding Virtual Memory article “The kscand task periodically sweeps through all the pages in memory, taking note of the amount of time the page has been in memory since it was last accessed. If kscand finds that a page has been accessed since it last visited the page, it increments the page’s age counter; otherwise, it decrements that counter. If kscand finds a page with its age counter at zero, it moves the page to the inactive dirty state.”

For the majority of Linux users and even system administrators on large servers this kernel process requires no intervention.  It is a simple process that works in the background doing its job well.  Nonetheless, under certain circumstances it can become necessary to tune kscand in order to improve system performance in a desirable way.

Issues with kscand are most likely to arise in a situation where a Linux box has an extremely large amount of memory and will be even more noticeable on boxes with slower memory.  The most notable is probably the HP Proliant DL585 G1 which can support 128GB or memory but in doing so drops bandwidth to a paltry 266MHz.  I first came across this particular issue on a server with 32GB of memory with approximately 31.5GB of it in use.  No swap space was being used and most of the memory was being used for cache so there was no strain on the memory system but the total amount of memory being scanned by the kscand process is where the issue truly lies.

Even on a busy server with gobs of memory (that’s the technical term) it would be extremely rare that kscand would cause any issues.  It is a very light process that runs quite quickly.  You are most likely to see kscand as a culprit when investigating problems with latency sensitive applications on memory intensive servers.  The first time that I came across the need to tune kscand was while diagnosing a strange latency pattern of network traffic going to a high-performance messaging bus.  The latency was minor but small spikes were causing concern in the very sensitive environment.  kscand was spotted as the only questionable process receiving much system attention during the high latency periods.

Under normal conditions, that is default tuning, kscand will run every thirty seconds and will scan 100% of the system memory looking for memory pages that can be freed.  This sweep is quick but can easily cause measurable system latency if you look carefully.  Through carefull tuning we can reduce the latency caused by this process but we do so as a tradeoff with memory utilization efficiency.  If you have a box with significantly extra memory or extremely static memory, such as large cache sizes that change very slowly, you can safely tune away from memory efficiency towards low latency with nominal pain and good results.

kscand is controlled by the proc filesystem with just the single setting of  /proc/sys/vm/kscand_work_percent. Like any kernel setting this can be changed on the fly to a live system (be careful) or can be set to persist through reboots by adding it to your /etc/sysctl.conf file.  Before we make any permanent changes we will want to do some testing.  This kernel parameter tells kscand what percentage of the system memory to scan each time that a memory scan is performed.  Since it is normally set to 100 kscand normally scans all in-use memory each time that it is called.  You can verify you current setting quite easily.

cat /proc/sys/vm/kscand_work_percent

A good starting point with kscand_work_percent is to set to 50.  A very small adjustment may not be noticeable so seeing 100 and then 50 should provide a good starting point for evaluating the changes in system performance.  It is not recommended to set kscand_work_percent below 10 and I would be quite wary of dropping even below 20 unless you truly have a tremendous amount of unused memory and your usage is quite static.

echo 50 > /proc/sys/vm/kscand_work_percent

Once you have determined the best balance of latency and memory utilization that makes sense for your environment you can make you changes permanent.  Be sure to only use the echo technique if this is the first time that this will be added to the file. You will need to edit it by hand after that.

echo "kscand_work_percent = 50" >> /etc/sysctl.conf

Keep in mind that the need to edit this particular kernel parameter is extremely uncommon and will need to be done only under extraordinary circumstances.  You will not need to do this in normal, everyday Linux life and even a senior Linux administrator could easily never have need to modify this setting.  On very specific conditions will cause this performance characteristic to be measurable or its modification to be desirable.

All of my testing was done on Red Hat Enterprise Linux 3 Update 6.  This parameter is the same across many versions although the performance characteristics of kscand vary between kernel revisions so do not assume that the need to modify the parameters in one situation will mean that it is needed in another.

RHEL 3 prior to update 6 had a much less efficient kscand process and much greater benefit is likely to be found moving to a later 2.4 family kernel revision.  RHEL 4 and later, on the 2.6 series kernels, is completely different and the latency issues are, I believe, less pronounced.  In my own testing the same application on the same servers moving from RHEL 3 U6 to RHEL 4.5 removed all need for this tweak even under identical load.  [Edit – In RHEL 4 and later (kernel series 2.6) the kscand process has been removed and replaced with kswapd and pdflush.]

Things that are likely to impact the behavior of kscand that you should consider include the following:

  • Total Used Memory Size, regardless of total available memory size.  The more you have the more kscand will impact you.  Determined by: free -m | grep Mem | awk '{print $3}'
  • Memory Latency, check with your memory hardware vendor. Higher latency will cause kscand to have a larger impact.
  • Memory Bandwidth.  Currently in speeds ranging from 266MHz to 1066MHz.  The slower the memory the more likely a scan will impact you and tuning will be useful.
  • Value in kscand_work_percent. The lower the value the lower the latency.  The higher the value the better the memory utilization.
  • Memory Access Hops.  Number of system bus hops necessary to access memory resources.  For example a two socket AMD Opteron server (HP Proliant DL385) never has more than one hop.  But a four socket AMD Opteron server (HPProliant DL585) can have two hops increasing effective memory latency. So a DL585 is more likely to be affected than a DL385 with all other factors being equal (as long as all three or four processor sockets are occupied.)

Join the Conversation


  1. It seems that kscand is removed from the newer linux kernel. I have several linux machines with 2.6.9 or 2.6.18 kernels. I can not find kscand running. The systems were left idle for several hours, but the active metric in meminfo does not change at all. I dont see a kscand_work_percent in /proc/sys/vm folder.

  2. You are correct. You will not see kscand in the 2.6.9 or 2.6.18 (RHEL 4 and RHEL 5) systems. kscand was removed with the 2.6 kernel series. Similar functionality will be found split between kswapd and pdflush processes in these more modern systems.

  3. It was a simple and thorough explanation of kscand process. We had to reduce the work_percent param in our RHEL 3 system for fixing node reboots after Oracle CRS upgrade. our case of because of high latency periods when kscand was running (It was running at 100%).

  4. Thank you for this nice article on kscand.

    One thing you mentioned is that with RHEL 4 and later, kscand has been replaced with kswapd. However, on RHEL 2.4.21-47, I see both kscand and kswapd running.

  5. RHEL 4 and later, which use a 2.6 kernel, do not have kscand. On your system which is older than RHEL 4 (2.4 kernels end with RHEL 3) you will see both kscand and kswapd. RHEL is not versioned 2.4.x, that is a kernel version. Even if 2.4 was a RHEL version, it would still be older than 4.

Leave a comment