On-demand accounting

User Beancounters Definition /proc/user_beancounters /proc/bc/ General information Units of measurement VSwap Parameters description Primary parameters numproc, numtcpsock, numothersock, vmguarpages Secondary parameters kmemsize, tcpsndbuf, tcprcvbuf, othersockbuf, dgramrcvbuf, oomguarpages, privvmpages Auxiliary parameters lockedpages, shmpages, physpages, numfile, numflock, numpty, numsiginfo, dcachesize, numiptent, swappages Internals User pages accounting RSS fractions accounting On-demand accounting UBC consistency Consistency formulae System-wide configuration vzubc(8) Configuration examples Basic Derived Intermediate configurations Tables List of parameters Parameter properties Consistency Config examples

This page describes a very promising way of beancounters optimization.

Current accounting model

Basically allocation of any kind of resource looks like this:

struct some_resource *get_the_resource(int amount)
{
        struct some_resource *ret;

        ret = find_or_allocate_the_resource(amount);
        return ret;
}

We change this behaviour to work like this:

struct some_resource *get_the_resource(int amount)
{
        struct some_resource *ret;

        if (charge_beancounter(amount) < 0)       
                return NULL;   

        ret = find_or_allocate_the_resource(amount);
        if (ret != NULL)
                return ret;

        uncharge_beancounter(amount);
        return NULL;
}

The charge_beancounter() call is responsible for checking whether the user is allowed to get the requested amount of the resource, i.e. if the resource consumption level is lower than the limit set.

Obviously, this change slows down the original code, as charge_beancounter() takes some slow operations like taking locks. We have an idea of how to optimize this behavior.

On-demand accounting basics

The main idea sonds like this:

If the consumption level of any resource can be easily upper estimated with some value, and this estimation is lower than the limit, then we do not need to know the exact consumption level and allow the resource allocation without additional checks

Apparently, when the estimation exceeds the limit, we must switch to the slower mode, that will give us more precise value of the consumption level and (probably) allocate another portion of the resource.

Example

Let's look at example of how this will work with the user memory accounting.

Currently we account for the physpages resource, that is, the number of physical pages consumed by a set of processes. The accounting hooks are placed inside the page fault handlers and thus hurting the performance. Currently accounting looks like this:

struct page *get_new_page(struct mm_struct *mm)
{
        struct page *pg;

        if (charge_beancounter(1) < 0)
                return NULL;

        pg = alloc_new_page(mm);
        if (pg != NULL)
                return pg;

        uncharge_beancounter(1);
        return NULL;
}

However, we have a good upper estimation of the RSS size — that is the lengths of mappings of the processes. Since the physical pages can only be allocated within these mappings, the RSS value can never exceed the sum of their lengths. The accounting will then look like this:

struct vm_area_struct *get_new_mapping(struct mm_struct *mm,
                unsigned long pages)
{
        if (!mm->fast_accounting)
                goto allocate;

        if (charge_beancounter(pages) == 0)
                goto allocate;

        mm->fast_accounting = 0;
        recalculate_the_rss(mm);

allocate:
        expand_mapping(mm);
}

struct page *get_new_page(struct mm_struct *mm)
{
        if (mm->fast_accounting)
                goto fast_path;

        if (charge_beancounter(1) < 0)
                return NULL;

fast_path:
        pg = alloc_new_page(mm);
        if (pg != NULL)
                return pg;

        if (!mm->fast_accounting)
                uncharge_beancounter(1);
        return NULL;
}

We do not call the slow charge_beancounter() function in the page fault (get_new_page()). Instead we account for the upper estimation in get_new_mapping() call that happens rarely and thus do not affect the performance.

Note, that the recalculate_the_rss() is called to calculate the exact RSS value on the beancounter.

More things to do

In this model we switch from the fast acounting to the slow one. However, if the upper estimation becomes lower than the limit again we can switch back to the fast model. However, these switches are not very fast themselves, and being too frequent can hurt the performance instead of improving. So this would require some further investigations.