Everything You Should Know About Intel SGX Performance on Virtualized Systems

1. Summary

Motivation of this paper

• Motivation

  • No work has studies in detail the performance degradation caused by SGX in virtualized systems.

    • this paper presents for the first time a detailed performance analysis of Intel SGX in a virtualized system
    • compare with bare-metal system

  • identify several optimization strategies to improve performance of Intel SGX

SGX Evaluation

1. Basis

• Intel SGX in a virtualized system

  • For each VM started with SGX support, the hypervisor allocates a section of the EPC to use a virtual EPC.

    • For several VMs, the hardware EPC is partitioned and distributed to several VMs.
    • Using KVM in this paper

• Evaluation setting

  • For local machine, kvm-sgx and qemu-sgx
  • Ubuntu 16.04
  • For VM, it uses 32MB of EPC capacity.

2. ECall/OCall Performance

• buffer in, out, in/out, user_check (zero-copy data transfer)

  • OCalls are generally faster than ECalls (10.6%-12.4)

3. Memory access

• write/read from the enclave vs. write/read from outside the application

  • outside application: write/read unencrypted memory
  • enclave: write/read unencrypted memory, write/read encrypted memory

• Result:

  • memory reads and writes into encrypted memory from the enclave is similar to performance without SGX
  • read/write into unencrypted memory from enclave is slightly slower (17%).

4. Swapping

• Evict process:

  • mark the given pages
  • no new TLB entries pointing to the page can be created
  • keep track of TLB entries concerning the page
  • OS must cause all processors executing inside the enclave to exit to flush these TLB entries
  • once all processors have exited from the enclave, the OS can write the sealed memory page to RAM

• Paging Measurement

  • multiple threads, accessing random addresses in a large memory buffer inside the enclave.

5. Enclave initialization and destruction

• Initialization: Before an enclave is ready to use, its memory contents are measured by CPU to produce a cryptographic hash

  • add a large static array into the enclave code and measured the enclave startup time

    • startup time vs. array size (varying from 0MB to 160MB)
    • it is measured by RDTSCP instruction.

• Destruction: large EPC also means more pages in the enclave's working set must be deallocated

  • leading to a slower shutdown

6. SGX-SSL benchmarks

• Measure the performance of encrypting memory buffers of various sizes from 16 bytes to 16MB using AES-128-GCM.

  • virtualized enclaves have a small (less than 10%) performance penalty over bare-metal enclaves.
  • the slowdown disappears with large buffer sizes.

2. Strength (Contributions of the paper)

1. do the benchmark on SGX in both bare-metal and virtualization machines.
2. also propose a non-intrusive method for measuring the characteristics of a given workload to identify whether it is suitable with nested paging or shadow paging.

3. Weakness (Limitations of the paper)

4. Some Insights (Future work)

1. Techniques on trusted computing

• Trusted platform module (TPM): used for platform integrity purpose by measuring critical system software (firmware, boot loaders, kernel, etc.) through the use of platform configuration registers

  • is not complete solution to trustworthy computing since the user still needs to trust many other components (e.g., OS, storage devices).

• Intel SGX: users no longer need to trust the system software nor any other hardware components. (self-contained solution)

• Virtualization-friendly trusted computing

  • AMD secure encrypted virtualization (SEV): a memory encryption-based technique:

    • protect virtual machines (VMs) from hypervisor attacks and hardware snooping.

  • Microsoft shielded VM

    • special VMs running on strictly controlled environments (e.g., guarded fabrics)
    • protected from snooping and alteration by the host administrator

2. SGX initialization

• At the boot time, the BIOS verifies whether SGX is enabled. It then reserves a region of physical memory for the CPU

  • designated as the enclave page cache (EPC).