Energy Estimation for Virtual Servers Using Resource Utilization and Machine Learning

Amandip Sangha

Fig. 1: Illustration of host–guest resource allocation in KVM virtualization and how host power interfaces such as RAPL are inaccessible to ordinary guests.

Fig. 2: Overview of the dual-logging setup: RAPL-based power logging on the host and resource utilization logging on the guest.

Set	Feature set	Features
F1	Basic	cpu.total mem.used diskio.vda.read_bytes_rate_per_sec diskio.vda.write_bytes_rate_per_sec network.enp1s0.bytes_recv_rate_per_sec network.enp1s0.bytes_sent_rate_per_sec
F2	Basic + CPU performance	F1 + perf.cycles_per_sec perf.instructions_per_sec perf.ref_cycles_per_sec perf.ipc
F3	Basic + memory performance	F1 + perf.dram_read_bytes_est_per_sec perf.l3_load_misses_per_sec perf.loads_per_sec
F4	Basic + I/O performance	F1 + perf.block_rq_issue_per_sec perf.block_rq_complete_per_sec perf.net_dev_xmit_per_sec perf.netif_receive_skb_per_sec
F5	All	Union of F1–F4

TABLE I: Guest-side feature sets used for model training.

Fig. 3: Representative within-platform prediction for kernel compilation on the Xeon 6527P. The upper panel compares measured and predicted VM-associated power over a held-out interval; the lower panel shows the signed prediction error,

\hat{Y}-Y

.

Platform	Workload	$\mathbf{R^{2}}$	MAE	RMSE	NMAE
			(W)	(W)	(%)
E3-1270 v6	kernelcomp	0.985	0.609	1.682	1.97
E3-1270 v6	apache	0.936	2.688	3.877	10.30
E3-1270 v6	pgbench	0.857	1.853	3.233	5.93
E3-1270 v6	iperf3	0.932	1.661	2.555	15.52
E-2286G	kernelcomp	0.961	2.450	3.900	5.58

TABLE II: Within-platform prediction performance using the full feature set.

Fig. 4: Mean grouped permutation-importance share across the five CPU platforms. The pooled column corresponds to the within-platform workload-pooled model on each hardware platform.

Fig. 5: Cross-platform variability of grouped permutation importance. Each cell gives the standard deviation, in percentage points, of the normalized resource-family importance across the five CPU platforms.

Platform	kernelcomp	Apache	pgbench	iperf3
E3-1270 v6	0.914	0.630	0.493	0.130
E-2286G	0.442	0.661	-1.153	-2.351
Xeon 6527P	0.756	0.738	0.751	0.208
EPYC 4245P	0.709	0.558	0.515	-1.395
EPYC 7443P	0.053	0.153	-1.331	0.516

TABLE III: Leave-one-workload-out

R^{2}

for each hardware platform and held-out workload.

Platform	kernelcomp	Apache	pgbench	iperf3
E3-1270 v6	-0.071	-0.306	-0.364	-0.802
E-2286G	-0.519	-0.284	-1.939	-3.267
Xeon 6527P	-0.196	-0.149	-0.169	-0.526
EPYC 4245P	-0.261	-0.401	-0.367	-2.250
EPYC 7443P	-0.914	-0.783	-2.266	-0.349

TABLE IV: Change in

R^{2}

under leave-one-workload-out generalization, relative to the corresponding within-workload

R^{2}

from RQ1:

\Delta R^{2}=R^{2}_{\mathrm{LOWO}}-R^{2}_{\mathrm{within}}

.

Fig. 6: Leave-one-workload-out prediction for held-out kernel compilation on the Xeon 6527P. The transferred model captures the principal temporal transitions but shows systematic underprediction in several higher-power regimes. The lower panel shows signed prediction error,

\hat{Y}-Y

.

Fig. 7: Leave-one-workload-out prediction for held-out iperf3 on the Xeon 6527P. In this weaker-transfer case, the model misses a large transient and under-represents later changes in power level, illustrating behavior not apparent from aggregate error metrics alone. The lower panel shows signed prediction error,

\hat{Y}-Y

.

Held-out platform	$\mathbf{R^{2}}$	MAE (W)	RMSE (W)	NMAE (%)
E3-1270 v6	0.566	7.614	10.087	31.62
E-2286G	0.468	14.526	18.151	40.51
Xeon 6527P	-0.373	7.737	10.914	39.53
EPYC 4245P	0.774	7.254	9.401	22.88
EPYC 7443P	0.625	5.761	7.114	25.61

TABLE V: Leave-one-hardware-out performance on each unseen hardware platform.

Fig. 8: Representative held-out prediction from the pooled multi-platform model. The upper panel compares measured and predicted VM-associated power; the lower panel shows the signed prediction error,

\hat{Y}-Y

.

Platform	$\mathbf{R^{2}}$	MAE (W)	RMSE (W)	NMAE (%)
All platforms	0.890	3.858	5.835	13.87
E3-1270 v6	0.889	3.496	4.966	14.16
E-2286G	0.862	6.495	9.027	17.68
Xeon 6527P	0.546	3.833	5.776	18.71
EPYC 4245P	0.952	2.675	4.000	8.02
EPYC 7443P	0.856	2.970	4.147	12.70

TABLE VI: Pooled multi-platform prediction performance, averaged over 100 temporally blocked evaluations.

Workload mix	$\mathbf{R^{2}_{\rm within}}$	$\mathbf{R^{2}_{\rm transfer}}$	$\mathbf{\Delta R^{2}}$	$\mathbf{NMAE}_{\mathrm{transfer}}$ (%)
Apache+pgbench	0.859	-0.178	-1.037	8.77
kernelcomp+Apache	0.951	0.845	-0.107	6.66
kernelcomp+iperf3	0.959	0.802	-0.157	11.12
kernelcomp+pgbench	0.756	0.487	-0.269	3.70

TABLE VII: Concurrent-workload prediction and transfer performance. Within and transfer metrics are averaged over the same 100 temporally blocked test intervals.

Configuration	$\mathbf{R^{2}_{\rm within}}$	$\mathbf{R^{2}_{\rm transfer}}$	$\mathbf{\Delta R^{2}}$	$\mathbf{R^{2}_{\rm pooled}}$
2 vCPU, 8 GB	0.935	0.453	-0.482	0.916
4 vCPU, 4 GB	0.969	0.747	-0.223	0.932
4 vCPU, 8 GB	0.973	–	–	0.960
4 vCPU, 16 GB	0.965	0.918	-0.047	0.943
8 vCPU, 8 GB	0.964	0.816	-0.148	0.957

TABLE VIII: Sensitivity to VM resource allocation on the EPYC 4345P. Within-configuration and pooled-model results are averaged over 100 temporally blocked evaluations. Transfer uses the 4-vCPU, 8-GB configuration as the reference.

Platform	$\mathbf{\Delta P}$ (W)	$\mathbf{\Delta E}$ (Wh)	Progress (%)	$E$ /action (%)
E3-1270 v6	0.071	0.0059	-4.00	4.41
E-2286G	0.097	0.0080	-4.23	4.55
Xeon 6527P	0.319	0.0266	-1.01	1.45
EPYC 4245P	0.326	0.0272	-1.10	1.60
EPYC 4345P	-0.744	-0.0620	-2.66	1.74
EPYC 7443P	-0.367	-0.0306	-6.25	6.34

TABLE IX: Monitoring overhead for Glances and perf on the six hardware platforms. Values are means over three matched ON/OFF pairs, with each run evaluated over the same 300-s interval.

Study	Prediction target	Guest-only inference	Runtime host / hypervisor access	Hardware diversity	Transfer / generalization	Concurrent / VM-config. evaluation
Joulemeter [10]	VM power	No	Yes	Multiple production servers	Not primary focus	No
VMeter [12]	VM and host power	No	Yes	Single experimental platform	Multiple workloads	No
CloudMonitor [13]	Server power	No	Calibration / provider-side	Hardware-specific calibration	Limited	No
Povoa et al. [14]	Whole-system power	No	OS/system access	Multiple hardware configurations	Cross-hardware modeling	No
Veni and Bhanu [16]	VM power	No	Performance-counter access	Limited	Multiple workloads	No
Present work	VM-associated CPU-package power	Yes, after training	Training only	Six Intel/AMD platforms	Workload and hardware transfer	Yes

TABLE S1: Comparison with representative prior approaches to server and VM-level power estimation.

Study	Input features / families	Reported prediction performance
Joulemeter [10]	CPU utilization and power states, memory activity, disk I/O; optional cache counters	Approximately 1–3 W error in reported multi-VM power-estimation experiments.
VMeter [12]	CPU, cache, disk, and DRAM resource-activity counters	93% mean and 94% median prediction accuracy against external power measurements.
CloudMonitor [13]	CPU, memory, disk, and network utilization	3.91% average prediction error on the evaluated server configuration.
Povoa et al. [14]	OS resource counters for CPU, memory, disk I/O, and network activity	Best reported result: 99.94% accuracy and 6.32 W error.
Veni and Bhanu [16]	Hardware PMCs (CPU cycles, DRAM accesses, instruction-cache fetches, LLC misses); disk I/O statistics	91–92% reported prediction accuracy using nonlinear SVR across benchmark workloads.
Present work	Guest CPU, memory, disk and network metrics; virtualized CPU and I/O performance counters	Pooled multi-platform: $R^{2}=0.890$ , MAE $=3.858$ W, NMAE $=13.87\%$ ; best within-platform: $R^{2}=0.985$ , MAE $=0.609$ W, NMAE $=1.97\%$ .

TABLE S2: Input features and reported prediction performance of representative prior approaches. Results are reproduced from the respective studies and are not directly comparable because prediction targets, hardware, workloads, and evaluation protocols differ.

Attribute	Intel Xeon E3-1270 v6	Intel Xeon E-2286G	Intel Xeon 6527P	AMD EPYC 4245P	AMD EPYC 4345P	AMD EPYC 7443P
Microarchitecture / generation	Kaby Lake / Xeon E3-1200 v6	Coffee Lake / Xeon E-2200	Granite Rapids-SP / Xeon 6	Zen 5 / EPYC 4005	Zen 5 / EPYC 4005	Zen 3 / EPYC 7003 (Milan)
Sockets	1	1	1	1	1	1
Cores / threads	4 / 8	6 / 12	24 / 48	6 / 12	8 / 16	24 / 48
Base / max frequency	3.8 / 4.2 GHz	4.0 / 4.9 GHz	3.0 / 4.2 GHz	3.9 / 5.484 GHz	3.8 / 5.581 GHz	2.8 GHz
Host memory	32 GB DDR4-2400 ECC, 2 $\times$ 16 GB Micron	32 GB DDR4-2667 ECC, 2 $\times$ 16 GB Micron	256 GB DDR5-6400 ECC, 8 $\times$ 32 GB Micron	32 GB DDR5-5600 ECC, 2 $\times$ 16 GB Micron	64 GB DDR5-5200 ECC, 2 $\times$ 32 GB Micron	256 GB DDR4-3200 ECC, 8 $\times$ 32 GB
Storage	2 $\times$ 240 GB SATA SSD	2 $\times$ 960 GB SATA SSD	2 $\times$ 1.92 TB NVMe SSD	2 $\times$ 960 GB NVMe SSD	2 $\times$ 1.92 TB NVMe SSD	2 $\times$ 480 GB SAS SSD

TABLE S3: Hardware and software characteristics of the physical host platforms.

Set	Model feature	Source	Meaning / construction	Unit
F1	cpu.total	Glances	Total CPU utilization observed inside the guest VM.	%
F1	mem.used	Glances	Guest memory currently in use.	bytes
F1	diskio.vda.read_bytes_rate_per_sec	Glances	Read throughput of the guest virtual block device vda.	bytes/s
F1	diskio.vda.write_bytes_rate_per_sec	Glances	Write throughput of the guest virtual block device vda.	bytes/s
F1	network.enp1s0.bytes_recv_rate_per_sec	Glances	Guest network receive throughput on interface enp1s0.	bytes/s
F1	network.enp1s0.bytes_sent_rate_per_sec	Glances	Guest network transmit throughput on interface enp1s0.	bytes/s

TABLE S4: Guest-side model features and their definitions. Rate-based perf features are normalized by the actual elapsed sampling interval.

Platform	Derived model feature	Raw perf event(s)	Construction / interpretation
Intel	perf.dram_read_bytes_est_per_sec	offcore_requests.l3_miss_demand_data_rd	Demand-data reads that miss the LLC are converted to a rate and multiplied by 64 bytes per cache line to obtain an estimated DRAM-read traffic rate.
Intel	perf.l3_load_misses_per_sec	mem_load_retired.l3_miss	Retired load instructions missing the L3 cache, divided by the measured perf interval.
Intel	perf.loads_per_sec	mem_inst_retired.all_loads	Retired load instructions divided by the measured perf interval.
AMD EPYC 4245P / 4345P	perf.sys_dram_io_fills_per_sec	ls_dmnd_fills_from_sys.dram_io_all	System demand fills serviced through DRAM I/O, divided by the measured perf interval.
AMD EPYC 7443P	perf.sys_dram_io_fills_per_sec	ls_dmnd_fills_from_sys.mem_io_local + ls_dmnd_fills_from_sys.mem_io_remote	Local and remote system-memory demand fills are summed and divided by the measured perf interval to form the same AMD derived feature.
AMD	perf.loads_dispatched_per_sec	ls_dispatch.ld_dispatch	Dispatched load operations divided by the measured perf interval.

TABLE S5: Architecture-specific memory-performance features used in feature set F3. Intel and AMD PMU events are retained with distinct semantics rather than treated as equivalent hardware counters.

Attribute	Configuration
Guest OS	Ubuntu 26.04 LTS
Kernel	Linux 7.0.0
Architecture	x86-64
Virtualization	KVM
Virtual hardware model	QEMU Q35 / ICH9
Virtual disk	100 GB

TABLE S6: Common guest-VM software and virtual-hardware configuration.

Fig. S1: Autocorrelation (

\rho

) of the VM-attributed power target at 1 Hz sampling. The dashed horizontal lines indicate

|\rho|=0.1

.

Fig. S2: Autocorrelation (

\rho

) of representative Glances predictors at 1 Hz sampling. The dashed horizontal lines indicate

|\rho|=0.1

.

Fig. S3: Autocorrelation (

\rho

) of representative perf predictors at 1 Hz sampling. The dashed horizontal lines indicate

|\rho|=0.1

.

Fig. S4: Grouped permutation importance for the Intel Xeon E3-1270 v6. Values show the decrease in held-out

R^{2}