Coalescent Computing Lit Review

last updated: 2026-05-19 01:18:15 -0700
commit: 9a29cb9
finished coalescent and moved old notes over

General Notes

Core Idea

Coalescent computing = cyber foraging for disaggregated edge hardware

Coalescence principle

Resources coalesce into a user’s device proportionally to physical proximity

Users’ devices appear to gain/lose compute, memory, GPU, etc. transparently as they move through an environment

Distinguished from typical cloud offload butomatic, proximity driven, fine grained

Hardware

Wireless latency is largest blocker (WiFi at ~20ms), but 5G URLLC can get to ~1ms

Cache coherence isn’t possible at these latencies

Current NICs not built for RDMA-style disagg

Need hardware extensions

acpi needs extensions to reflect dynamically coalesced resouces

Software

Need a “Coalescent OS” supporting:

Performance, disaggregation, resource discovery, adaptation, heterogeneity, fault tolerance
LegoOS is closest foundation
- Designed for datacenter so ExCache and Infiniband based RPC don’t translate

Scheduling is multi-objective optimization problem that is further complicated by user mobility

Centralized won’t work, so use decentralized ML approach

Heterogeneous ISAs require fat binaries or a JIT based IR approach

Resource discovery needs to balance reactivity with network and battery overhead

Programming Model

Default: OS transparently migrates threads to remote vCPUs

Programmer unaware

Explicit API: cthread wrapper lets programmer specify latency bounds, remote preference, failure handlers

FaaS variant: Virtines used as primitive for synchronous remote function invocation, in connection to Wanninger et al. (2022)

Fault Tolerance

Replica hierarchy

Edge server (secondary) \rightarrow cloud (tertiary, relaxed consistency)

Periodic checkpointing + append-only storage for state recovery

Best-effort coalescence

If network can’t meet SLO, fall back to local or cloud

Key Findings

Coalescent Computing is positioned as feasible in the near term given converging trends: disaggregated hardware, hierarchical edge clouds, and ultra-low-latency wireless (5G URLLC, WiFi 6E)

Virtines are the proposed primitive for the FaaS/RPC programming model in a coalescent system

Deterministic performance is identified as the central unsolved problem

Same albatross that killed early distributed OSes

References (if any)

Wanninger, Nicholas C., Joshua J. Bowden, Kirtankumar Shetty, Ayush Garg, and Kyle C. Hale. 2022. “Isolating Functions at the Hardware Limit with Virtines.” Proceedings of the Seventeenth European Conference on Computer Systems, EuroSys ’22, March 2022, 644–62. https://doi.org/10.1145/3492321.3519553.

diff --git a/lit-reviews/hale2021coalescent.md b/lit-reviews/hale2021coalescent.md index e8066f1..6f8fcdc 100644 --- a/lit-reviews/hale2021coalescent.md +++ b/lit-reviews/hale2021coalescent.md @@ -5,3 +5,90 @@ date: "2026-05-19" bibliography: bibliography/references.bib csl: bibliography/chicago-author-date.csl --- + +## Source Information + +- **Title:** Coalescent Computing +- **Authors:** Kyle C. Hale +- **Year:** 2021 +- **Publication:** APSys 2021 +- **PDF:** [Coalescent Computing](/assets/papers/hale2021coalescent) + +## General Notes + +### Core Idea + +- Coalescent computing = cyber foraging for disaggregated edge hardware +- Coalescence principle + - Resources coalesce into a user's device proportionally to physical proximity +- Users' devices appear to gain/lose compute, memory, GPU, etc. transparently as + they move through an environment +- Distinguished from typical cloud offload b\c automatic, proximity driven, fine + grained + +### Hardware + +- Wireless latency is largest blocker (WiFi at ~20ms), but 5G URLLC can get to + ~1ms + - Cache coherence isn't possible at these latencies +- Current NICs not built for RDMA-style disagg + - Need hardware extensions +- [[../concepts/acpi]] needs extensions to reflect dynamically coalesced + resouces + +### Software + +- Need a "Coalescent OS" supporting: + - Performance, disaggregation, resource discovery, adaptation, heterogeneity, + fault tolerance + - LegoOS is closest foundation + - Designed for datacenter so ExCache and Infiniband based RPC don't + translate +- Scheduling is multi-objective optimization problem that is further complicated + by user mobility + - Centralized won't work, so use decentralized ML approach +- Heterogeneous ISAs require fat binaries or a JIT based IR approach +- Resource discovery needs to balance reactivity with network and battery + overhead + +### Programming Model + +- Default: OS transparently migrates threads to remote vCPUs + - Programmer unaware +- Explicit API: `cthread` wrapper lets programmer specify latency bounds, remote + preference, failure handlers +- FaaS variant: [[../concepts/virtines]] used as primitive for synchronous + remote function invocation, in connection to @wanninger2022isolating + +### Fault Tolerance + +- Replica hierarchy + - Edge server (secondary) $\rightarrow$ cloud (tertiary, relaxed consistency) +- Periodic checkpointing + append-only storage for state recovery +- Best-effort coalescence + - If network can't meet SLO, fall back to local or cloud + +## Key Findings + +- Coalescent Computing is positioned as feasible in the near term given + converging trends: disaggregated hardware, hierarchical edge clouds, and + ultra-low-latency wireless (5G URLLC, WiFi 6E) +- [[../concepts/virtines]] are the proposed primitive for the FaaS/RPC + programming model in a coalescent system +- Deterministic performance is identified as the central unsolved problem + - Same albatross that killed early distributed OSes + +## Critique/Gaps + +- This is a concept paper, so there's a lot lol + +## Questions + +- Is a trust peer and/or participation peer a good idea? + +## Relations + +- [[../concepts/unikernels]] +- [[../concepts/disaggregated-hardware]] + +## References (if any)