Integrated Performance Score (IPS)

• Integrated Performance Score (IPS) is a target-based metric that aggregates normalized resource utilizations (CPU, memory, I/O, etc.) to assess a machine's overall health.
• It employs nonlinear mappings like tanh and exponential decay to quantify deviations from ideal targets, facilitating balanced resource management.
• The framework uses customizable weights, tolerances, and penalty terms to support precise capacity planning and effective VM right-sizing in cloud environments.

Searching arXiv for the specified paper and closely related context.
The Integrated Performance Score (IPS), as instantiated by the WISE framework (“Workload/Machine Index Score”), is a single-value fitness indicator for a given workload–machine combination. Its purpose is to answer whether CPU, memory, I/O, and related resources are running too low, too high, or “just right,” how far each resource’s utilization deviates from its pre-defined ideal target, and what the aggregated health of the machine is under the workload. By mapping each resource’s utilization to a normalized score, applying weights, and penalizing any extreme over-utilization, IPS provides a rapid, tunable guide for capacity planning, right-sizing VMs, and automated instance selection [2012.07984].

1. Definition and operational objective

IPS is defined as a single-score summary of the fitness of a computing machine under a specified workload configuration. In the WISE formulation, the score is explicitly target-based rather than purely utilization-based: it evaluates how far machine resources are operating from target levels and whether those resources are being stressed or sitting idle wasting precious resources.

The framework is designed to encompass any number of computing resources. The examples and notation include CPU, memory, disk I/O, and network, but the formulation is not limited to those resources. The central operational question is not merely whether a machine is busy, but whether the observed resource profile is balanced relative to pre-defined ideal targets and upper safety limits.

This target-centered design gives IPS a dual role. First, it is an aggregate health measure for a workload–machine combination. Second, it functions as a decision aid for resource allocation, because it distinguishes between under-utilization, over-utilization, and near-target operation. A plausible implication is that IPS is intended for comparative evaluation across machine instances only after the target, tolerance, weighting, and penalty definitions have been fixed for the workload of interest.

2. Mathematical structure

The WISE instantiation of IPS introduces the following notation [2012.07984]:

Symbol	Meaning
$n$	number of monitored resources
$u_i$	actual utilization rate of resource $i$
$t_i$	target (ideal) utilization for resource $i$
$\sigma_i$	acceptable deviation (tolerance) around $t_i$
$r^{\max}_i$	maximum safe utilization for resource $i$
$w_i \ge 0$	relative importance weight of resource $i$
$\alpha \ge 0$	penalty-scaling factor

The first step is a standardized deviation from target:

$$
z_i \;=\;\frac{u_i - t_i}{\sigma_i}\,.
$$

A simple distance function is

$$
d_i \;=\;|z_i|\,,
$$

though WISE also uses nonlinear mappings of $z_i$.

A binary penalty is imposed if a resource exceeds its maximum safe level:

$$
P_i \;=\;\alpha \cdot H\bigl(u_i - r^{\max}_i\bigr),
\quad H(x)=
\begin{cases}
1& x\ge0,\
0& x<0.
\end{cases}
$$

At the resource level, WISE specifies two common normalized mappings. The hyperbolic-tangent variant has range $-1$ to $+1$:

$$
s_i^{\tanh} = \tanh(z_i)\,.
$$

Its interpretation is sign-sensitive: $s_i<0$ indicates under-utilized, $s_i>0$ indicates over-utilized, and $s_i=0$ indicates on target. The exponential-decay variant has range $0$ to $1$:

$$
s_i^{\exp} = \exp\bigl(-|z_i|\bigr)\,.
$$

Its interpretation is target-proximity: $s_i=1$ indicates exactly on target, while $s_i \to 0$ indicates far from target.

The integrated score admits four variants by combining either mapping with either an $\ell_1$- or $\ell_2$-norm aggregation. Two representative forms are:

1. Exponential + $\ell_1$ (best-is-highest, range $[0,1]$):
   $$
   S_{\rm IPS}
   \;=\;
   \max!\biggl[\,\frac{1}{n}\sum_{i=1}^n w_i\,\exp\bigl(-|z_i|\bigr)\;-\;\sum_{i=1}^n P_i\,,\;0\biggr],
   $$
   where a value of $1$ is ideal and $0$ is worst.

2. Tanh + $\ell_2$ (best-is-lowest, range $[0,1]$):
   $$
   S_{\rm IPS}
   \;=\;
   \min!\Bigl[\,\frac{1}{n}\,\sqrt{\sum_{i=1}^n\bigl(w_i\,\tanh(z_i)\bigr)^2}\;+\;\sum_{i=1}^n P_i\,,\;1\Bigr],
   $$
   where $0$ is ideal and $1$ is worst.

The more general statement is

$$

S_{\rm IPS}

\begin{cases}
\displaystyle
\max\Bigl[
\frac1n\sum_i w_i\,f\bigl(|z_i|\bigr)\;-\;\sum_i P_i,\,0
\Bigr], & \text{(exp-type)},\[6pt]
\displaystyle
\min\Bigl[
\frac1n\bigl|\,w\odot f(z)\bigr|_2\;+\;\sum_i P_i,\,1
\Bigr], & \text{(tanh-type)},
\end{cases}
$$

where $f(z)=\exp(-z)$ or $f(z)=\tanh(z)$, as required.

3. Computation procedure

The IPS workflow is given as a seven-step algorithm for a workload and a machine instance [2012.07984].

First, metric collection samples each resource’s utilization $u_i$ over the workload run, with examples including CPU avg, p95; RAM avg, disk I/O, and network. Second, targets and limits are defined: for each resource $i$, one sets an ideal target $t_i$, a tolerance $\sigma_i$, and an upper limit $r^{\max}_i$ triggering hard penalty. Third, the framework computes standardized deviations as $z_i = (u_i - t_i)/\sigma_i$ for $i=1\ldots n$.

Fourth, resource scores are computed by choosing either $s_i = \tanh(z_i)$ or $s_i = \exp(-|z_i|)$, with optional multiplication by weight $w_i$. Fifth, penalties are computed as $P_i = \alpha \cdot H(u_i-r^{\max}_i)$. Sixth, the selected aggregation formula is applied, using either the $\ell_1$- or $\ell_2$-based IPS expression together with max/min clamping. Seventh, the resulting score is interpreted relative to the corresponding best extreme.

The interpretation step is asymmetric across score families. If IPS is near its “best” extreme, namely $0$ for tanh-type or $1$ for exp-type, the machine is well-balanced. Scores beyond pre-set thresholds, exemplified by IPS $< 0.36$ or IPS $> 0.76$, signal under/over-utilization. This suggests that IPS is not a single canonical scalar in isolation; its operational meaning depends on the selected mapping, norm, and threshold policy.

4. Illustrative example

The WISE presentation includes a three-resource example with equal weights $w_i=1$ and no $\alpha$-scaling, $\alpha=1$ [2012.07984]. The parameters and observations are:

Resource	$t_i$, $\sigma_i$, $r^{\max}_i$	Observed $u_i$
CPU avg	$40\%$, $30\%$, N/A	$70\%$
RAM avg	$50\%$, $20\%$, $90\%$	$80\%$
Disk avg	$60\%$, $20\%$, N/A	$60\%$

The standardized deviations are

$$
z_{\rm CPU} = (0.70-0.40)/0.30 = 1.0
$$

$$
z_{\rm RAM} = (0.80-0.50)/0.20 = 1.5
$$

$$
z_{\rm Disk} = (0.60-0.60)/0.20 = 0.0\,.
$$

No resource exceeds $r^{\max}$, so all $P_i=0$.

Using the exp-variant + $\ell_1$ aggregation,

$$
s_{\rm CPU} = e^{-|1.0|}=0.3679
$$

$$
s_{\rm RAM} = e^{-1.5}=0.2231
$$

$$
s_{\rm Disk} = e^{-0}=1.0000
$$

$$
\mathrm{IPS} = (0.3679 + 0.2231 + 1.0000)/3 - 0 = 0.5303\,.
$$

The stated interpretation is that IPS $\approx 0.53$ on $[0,1]$ is moderately healthy but memory is somewhat over-utilized.

Using the tanh-variant + $\ell_2$ aggregation,

$$
s_{\rm CPU} = \tanh(1.0)=0.7616
$$

$$
s_{\rm RAM} = \tanh(1.5)=0.9051
$$

$$
s_{\rm Disk} = \tanh(0)=0.0
$$

$$
\ell_2\text{-norm} = \sqrt{(0.7616)^2+(0.9051)^2+0^2}/3 = \sqrt{1.38}/3 = 1.175/3 = 0.3917
$$

$$
\mathrm{IPS} = \min(0.3917 + 0,\,1) = 0.3917\,.
$$

The stated interpretation is that IPS $\approx 0.39$ on $[0,1]$, close to ideal $(0)$, indicates small deviations. The example shows that the same utilization vector can yield distinct numerical semantics depending on whether the scoring family is best-is-highest or best-is-lowest.

5. Use in resource allocation and right-sizing

IPS is presented as a guide for resource allocation decisions because it links aggregate scoring to concrete resizing actions. If IPS stays within a “sweet-spot” band, exemplified by $>0.36$ for exp-type or $<0.76$ for tanh-type, the instance is well-balanced [2012.07984].

Scores below the exp-type threshold or above the tanh-type threshold suggest re-sizing. The specified actions are: under-utilization implies considering downsizing to save cost, while over-utilization implies considering upsizing or adding capacity to avoid performance degradation. Per-resource $s_i$ diagnostics are used to pinpoint which resource drives the score.

This use of IPS is structurally different from thresholding raw utilization streams independently. Because the score is target-based and weighted, a machine can be judged suboptimal even when no single resource is near saturation, provided the overall operating point is far from the intended target profile. Conversely, the penalty term reserves special treatment for extreme over-utilization by explicitly checking $u_i > r^{\max}_i$.

A plausible implication is that IPS can support both human-in-the-loop capacity planning and automated instance selection, provided the target and tolerance definitions capture the workload’s desired operating regime.

6. Relation to conventional metrics and experimental validation

Conventional monitoring often reports raw CPU%, RAM%, or throughput in isolation. IPS offers multi-resource aggregation, target-based normalization, nonlinear mapping, and customizability [2012.07984]. More specifically, it merges CPU, memory, disk, and network into one unified “health” measure; measures deviation from an ideal operating point rather than merely “busy vs. idle”; softens the impact of extreme outliers while still flagging dangerous over-utilization via penalties; and supports per-resource weights, diverse aggregations such as avg and p95, and user-defined thresholds.

The experimental validation described for WISE uses AWS EC2 experiments with two distinct benchmarks: MongoDB as CPU-intensive and Streaming as network-intensive. In those experiments, the WISE-derived IPS was used to rank instance types. Validation against ground-truth performance metrics—latency, throughput, and cost—yielded precision of selecting truly optimal instances often 100%, recall up to 100% especially for exp + $\ell_2$ variant, and rank-biased overlap greater than $0.8$ between IPS ranking and performance-based ranking.

Those results are presented as evidence that IPS reliably identifies right-sized VMs and can drive automated instance recommendations in cloud environments. Within the WISE framework, the significance of these results lies in showing that a target-based aggregate resource score can align with workload-level performance rankings, rather than merely summarize low-level telemetry.